Latest from Tom's Hardware in Amd-ryzen-threadripper-1950x

MORE: All Motherboard Content

Test Configuration and Benchmark Results

Test System Components

RAM	(1) G.Skill Trident Z F4-3466C18Q-32GTZRXB 4x8GB DDR4-3466 CL18 (1) G.Skill Flare-X F4-3200C14Q-32GFX 4x8GB DDR4-3200 CL14 (1) Corsair Dominator Platinum CMD32GX4M4C3466C16 4x8GB DDR4-3466 CL16 (2) Viper Steel Series PVS416G373C7K 16GB (2x8GB) DDR4-3733 CL17
Sound	Integrated HD audio
Network	Integrated Gigabit networking
Chipset	AMD X399
Graphics Driver	NVIDIA GeForce 399.24
OS	Microsoft Windows 10 Anniversary Update (not 1903)

Our X399 test stand continues to sport its original launch Ryzen Threadripper 1950X processor and is cooled by the Fractal Design Celsius S36 AIO. We've thrown several memory kits at this board for our overclocking segment, but our main benchmarks continue to use the G.Skill Flare X kit listed above. Also, we continue to test our X399 boards with the previous Windows updates and do not want to muddy the water with the 1903 update.

Benchmark Settings

Synthetic Benchmarks and Settings
PCMark 8	Version 2.7.613Home, Creative, Work, Storage, Applications (Adobe & Microsoft)
SiSoftware Sandra	Version 2016.03.22.21CPU Arithmetic, Multimedia, Cryptography, Memory Bandwidth
IOmeter	k Random Read, 4k Random Write 128k Sequential Read, 128k Sequential Write
Cinebench R15	Build RC83328DEMOOpenGL Benchmark
CompuBench	Version 1.5.8Face Detection, Optical Flow, Ocean Surface, Ray Tracing
3D Tests and Settings
3DMark 13	Version 4.47.597.0 Test Set 1: Skydiver, 1920x1080, Default Preset Test Set 2: Firestrike, 1920x1080, Default Preset Test Set 3: Firestrike Extreme, 2560x1440 Default Preset
Application Tests and Settings
HandBrake CLI	Version: 0.9.9Sintel Open Movie Project: 4.19GB 4K mkv to x265 mp4
LAME MP3	Version 3.98.3Mixed 271MB WAV to mp3: Command: -b 160 --nores (160Kb/s)
Adobe After Effects CC	Release 2015.3.0, Version 13.8.0.144PCMark-driven routine
Adobe Photoshop CC	Release 2015.5.0, 20160603.r.88 x64PCMark-driven routine (light and heavy)
Adobe InDesign CC	Release 2015.4, Build 11.4.0.90 x64PCMark-driven routine
Adobe Illustrator	Release 2015.3.0, Version 20.0.0 (64-bit)PCMark-driven routine
Blender	Version 2.68aBMW 27 CPU Render Benchmark, BMW 27 GPU Render Benchmark
7-Zip	Version 16.02THG-Workload (7.6GB) to .7z, command line switches "a -t7z -r -m0=LZMA2 -mx=9"
Game Tests and Settings
*Ashes of the Singularity: Escalation*	Version 2.50.28527 High Preset - 1920x1080, Mid Shadow Quality, 1x MSAA Crazy Preset - 1920x1080, High Shadow Quality, 2x MSAA High Preset - 3840x2160, Mid Shadow Quality, 1x MSAA Crazy Preset -3840x2160, High Shadow Quality, 2x MSAA
*F1 2015*	2015 Season, Abu Dhabi Track, Rain 1920x1080 - UltraHigh Preset, 16x AF 3840x2160 - UltraHigh Preset, 16x AF
*Metro: Last Light Redux*	Version 3.00 x64 High Quality, 1920x1080, High Tesselation Very High Quality, 1920x1080, Very High Tesselation High Quality, 3840x2160, High Tesselation Very High Quality, 3840x2160, Very High Tesselation
*The Talos Principle*	Version 267252 1920x1080 - Medium Preset, High Quality, High Tesselation, 4x AF 1920x1080 - Ultra Preset, VeryHigh Quality, VeryHigh Tesselation, 16x AF 3840x2160 - Medium Preset, High Quality, High Tesselation, 4x AF 3840x2160 - Ultra Preset, VeryHigh Quality, VeryHigh Tesselation, 16x AF

Synthetics, Applications, and Gaming Benchmarks

The Asus Zenith Extreme Alpha is a thoroughbred when compared to our other offerings today but shows middle-of-the-pack performance in the PCMark basic tests. It outperforms the field with both the Adobe and Office suite. Sandra continues to show tight performance margins across the scenarios, though the Asus board takes the prize with memory bandwidth.

NVMe performance is a strong showing for the Alpha as well, and note this data is not from the DIMM.2 riser but from the onboard slot. Though losing in single-core performance in Cinebench R15, multi-core performance is excellent. Compubench shows mixed results across all platforms, adding confidence that we have not introduced any systematic errors at least into the graphical segment of our test bench.

Handbrake render times are impressive, showing a nearly 2% difference compared to the slower performers. Since the Asus board performed well in PCMark’s Adobe runs, digging further into the results shows hefty wins in both Photoshop Heavy, InDesign, and Illustrator (though After Effects does tend to struggle). Blender and Lame64 show slower times for the ROG, but 7-zip favors its memory bandwidth and shaves a few seconds compared to the Gigabyte Gaming 7.

For 3DMark, we see pretty noisy results across the samples with Skydiver, but moving on to Firestrike and Firestrike Extreme, performance stabilizes and ultimately favors no board.

Ashes of the Singularity shows similar behavior across the resolutions and graphical settings. Metro Last Light appears to favor the Asus Zenith and Gigabyte Gaming 7 boards across all presets and resolutions and all measurements land within 10 frames per second (fps) for all samples today.

The Talos Principle bucks the trend for the Asus board by favoring both Gigabyte boards, seeing a near 10 fps advantage across resolutions and settings. F1 brings our gaming segment to a close-by shocking us with the Aorus Gaming 7's result at 1080p, but it loses steam at higher settings and the Asus X399 Zenith Extreme Alpha takes the win for average frame rates at 4K.

Overall performance shows nearly perfect marks for the Asus X399 Zenith Extreme Alpha, but it slips up in applications by a hair. Interestingly, the Gigabyte Aorus X399 Xtreme manages to nearly hit that 100% performance mark in all respects and the Aorus Gaming 7 pulls off impressive gaming marks. ASRock's Cadillac offering, meanwhile, struggles in all but applications.

MORE: Best Motherboards

MORE: All Motherboard Content

Power, Thermals, and Overclocking

Threadripper is by no means an efficient processor line, nor are its platforms intended for high-efficiency deployments. The Asus Zenith Extreme Alpha manages to hang with the pack at idle, but once loads are applied the system consumes more energy. These numbers are particularly impacted by the GPU in this instance, but overall, results are commendable here. That said, on average, this system tends to burn nearly 30W more than competing products, which results in lower efficiency numbers.

Thermally speaking, the Zenith Extreme Alpha starts to show some benefits of its VReg design. While producing CPU temperatures on par with Gigabyte’s top-tier offering, the passive VReg cooler on that board just can’t stand up against the Alpha’s cooler and phase counts. We will talk about this more shortly, but this just goes to show what this board is capable of.

To do a true analysis of this board’s potential (and having access to more hardware), we want to throw all the hardware at we can at this board. Unfortunately, that takes time and coordination of hardware assets. As advertised, this board supports up to a DDR4-3600 data rate, and sure enough, we can hit that on several types of kits. First, we populated our standard Flare-X DDR4-3200 kit and were able to hit a respectable DDR4-3266 speed, which is good for this kit. Next, we swapped in the Trident-Z DDR4-3466 kit that we’ve recently reviewed and with minimal effort and minor voltage adjustments, we hit DDR4-3600. Not bad.

For grins, we swapped in our X470’s DDR4-3466 Dominator Platinums that also reached the advertised cap but required a few clicks of additional latency to get stable performance. Lastly, Viper sent us a couple 2x8GB kits of its DDR4-3733 RAM. Unfortunately, XMP fails to load and we must increase the voltage up to 1.4V to run at DDR4-3600. With all that said, the Asus Zenith Extreme Alpha manages to provide the foundation for impressive memory overclocks, although it's not a good testbed for continuous memory swapping.

Shifting gears back to the main reason this board is on the market, Asus has pulled all the stops for unleashing Threadripper. First off, the VReg on this system is a teamed eight-phase design which utilizes a total of sixteen MOSFET drivers. The advertised benefit to this is that the VReg circuit can handle voltage droop events better than conventional designs with up to 8.7% less voltage droop. To phrase it another way, this design should enable overclockers to not fear operating system scheduling issues and variations in the workload or implement less-aggressive load line settings.

With the addition of more power stages, Asus also implements an active cooling solution which only engages when the VRegs cross the 60 degree C threshold. This, in turn, increases the VReg’s efficiency window and allows for longer, higher-wattage CPU load scenarios. In our standard testing, this fan rarely ever turns on, which is proof that the extra surface area and efficiency of the heatsink are sufficient. Even when overclocking our 1950X, the VReg fans only occasionally engage, further proving the point that this design is intended for extreme overclocking. Had we utilized a chiller or other exotic cooling methods, there is probably over 50C of VReg thermal headroom to play with.

Needless to say, our standard overclocking test and results do not serve this product justice. We run into our Tdie thermal ceiling of 85C at a mild 3950MHz during an 8-hour Prime95 run and our Fractal Designs S36 just is not a good match for this motherboard. To put it in perspective, most other boards that run similar overclocks are typically starting to surpass the PROCHOT threshold for their voltage regulators. Though the chart suggests otherwise, full marks are rewarded to the Zenith Extreme Alpha.

MORE: Best Motherboards

MORE: All Motherboard Content

Value, Conclusions, and Going Forward

As we mentioned before, this thoroughbred Asus X399 Zenith Extreme Alpha is in a league of its own, and that pedigree comes at a cost. Even compared to the more well-equipped competitors (Gigabyte’s Xtreme and ASRock Professional Gaming), the almost USD $200 premium strips away the value consideration as we traditionally score it. Those dollars can go a long way in today’s market bumping builders up to higher-end GPUs, CPUs, additional NVMe, or other specialized components. Even if we want to talk cooling enhancements, dumping an additional 200 USD into blocks, fittings, and radiators can stretch a long way.

The experience we’ve had with the Asus X399 Zenith Extreme Alpha has been a great one, except for when swapping in different DIMMs for testing. As a builder, this system might be overkill, but for reviewers or professional overclockers, there is a reason the industry often gravitates towards Asus.

Back to Threadripper in 2019: Amazon’s Prime Day deals and the deep discounts we are seeing on 1st and 2nd generation Threadrippers lead us to believe there is now a value argument to be had in this HEDT platform. More importantly, content creators that were hesitant to invest in the platform now have options to increase their workflow and creative capacity. If we take Ryzen 3000 pricing into account, this $600 motherboard is on par with some of the competing “mainstream” X570 offerings. Would you rather buy a $500 Threadripper 2950X and $300 X399 motherboard, or $500 Ryzen 3900X and $300 X570 board? What makes sense will of course depend on what you want to do with your system and build. Let us know where you stand in the comment threads in the forums.

Image Credits: Tom's Hardware

MORE: Best Motherboards

MORE: All Motherboard Content

Gigabyte X399 Aorus Xtreme Review: Overclocking Without Compromise

Jacob Terkelsen — Tue, 15 Jan 2019 14:00:00 +0000

Intro and Product Description

Even with CES 2019 and more announcements from AMD stealing the spotlight, hardware reviewers are still in the trenches tinkering with hardware and learning more about different vendors’ implementations. Having given ASRock's motherboards a few turns recently on our test bench, it’s time give Gigabyte the spotlight. In short, the X399 Aorus Xtreme impressed us for several reasons, although the $425 price will certainly give some potential buyers pause.

Specifications

Socket	TR4
Chipset	AMD X399
Form Factor	E-ATX
Voltage Regulator	10+3 Phases
Video Ports	✗
USB Ports	10Gbps:(1) Type A, (1) Type C5Gb/s: (8) Type A
Network Jacks	(1) 10Gb Ethernet(2) Gigabit Ethernet
Audio Jacks	(5) Analog, (1) Digital Out
Legacy Ports/Jacks	✗
Other Ports/Jack	(2) SMA Antenna, (1) Power Button, (1) Clear CMOS
PCIe x16	(4) v3.0 (x16/x8/x16/x8)
PCIe x8	✗
PCIe x4	✗
PCIe x1	(1) v2.0
CrossFire/SLI	4x / 4x
DIMM slots	(8) DDR4
M.2 slots	(1) PCIe v3 x4 / SATA3 42/60/80mm(2) PCIe v3 x4 / SATA3 60/80/110mm
U.2 Ports	✗
SATA Ports	(6) 6Gb/s
USB Headers	(1) USB3.1 Gen2(2) 5Gbps(2) USB2.0
Fan Headers	(7) 4-pin
Legacy Interfaces	✗
Other Interfaces	(1) TPM, (2) Thermistor, (2) Addressable LED, (2) RGBW, (1) SPDIF, (1) Front panel audio, (1) Clear CMOS Jumper
Diagnostics Panel	Numeric
Internal Button/Switch	Dual BIOS switch, BIOS switch
SATA Controllers	Integrated (0/1/10)
Ethernet Controllers	(1) Aquantia® AQC107, (2) Intel® I211AT
Wi-Fi / Bluetooth	Intel® 802.11ac WiFi Module, Bluetooth 4.2 / 3.0
USB Controllers	ASMedia® USB 3.1 Gen 2 Controller
HD Audio Codec	ALC1220-VB
DDL/DTS Connect	✗ / ✗
Warranty	3 Years

Released in the third quarter of 2018, the Gigabyte X399 Aorus Xtreme takes a lot of lessons learned from the company's previous renditions of X399 and X370 boards and adapts its designs to better suit the 2000 series Threadripper CPUs. Primarily, improvements in the cooling, power delivery, and IO compatibility are in place to better harness the platform’s potential. However, with those improvements Gigabyte decided to increase the planar size to the E-ATX form factor--though at only 10.6" depth, it will also fit many ATX and all XL-ATX cases. Be sure to check your case specs before buying.

Like the X399 Designare EX, the contents of the Aorus Xtreme package delivers quality and quantity with six braided SATA3 cables, two thermistors, two RGB extenders, two digital LED adapters, an Allen wrench for the M.2 heatsinks, a T20 Torx wrench for the CPU socket, Wi-Fi antenna, G-Connector, and a single dual-spaced SLI bridge. Also included are various screws and standoffs for the M.2 drives, Velcro cable ties, documentation, and a driver disc.

Glancing at the integrated backplate, five gold-plated analog and one SPDIF port are available for the ESS Sabre HIFI audio DAC and the ALC1220-VB codec, which boasts improved Signal-to-noise ratios and audio fidelity. For networking connections, two gigabit RJ-45 jacks are available through the tried-and-true Intel NIC, WiFi is supported by a dual-band 802.11ac Intel NIC, and a red RJ-45 jack is powered via the Aquantia network controller for 10Gb networking. As for ports, back panel USB is all USB 3.0 or greater, sporting eight USB3.0 and 2 USB3.1 10Gbps via 1 Type-A and 1 Type-C. And for the bench-top testers out there, Power and Clear CMOS buttons are available at the top side of the plate.

Gigabyte pivots from their Designare EX and Aorus Gaming 7 X399 layouts toward alternating PCIe and M.2 slots, akin to some ASRock and MSI X399 boards we’ve previously reviewed. Walking down the PCI slots, we have a four x16 slots wired for x16, x8, x16, x8 at Gen3 speeds. Located directly in the middle of those slots is a PCIe x1 slot nestled directly behind an NVMe port, and equipping that slot will throw a wrench into dual-slot card installation at the PCIEX8_1 location. Bottom edge connectors provide front panel audio, two RGB LED headers, TPM, BIOS and slow boot switches, two USB 2.0 and two USB 3.0 headers, three 4-pin fans, numeric debug LED, and front panel chassis pins.

With the added E-ATX real estate, the right edge surprisingly doesn’t add much extra connectivity than previous renditions. Six angled SATA3 ports are located on the bottom-right edge, along with an auxiliary PCIe power header for multi-GPU card configurations. Front panel USB 3.1 Gen2 is a nice addition and is located below the 24-pin ATX plug, and three more four-pin fan headers surround the two additional RGB headers. The top of the board gains a little bit of breathing room thanks to the larger board, and we only see additional connectors towards the rear IO with two 8-pin EPS connectors and our last of seven 4-pin fan headers. In case you were curious, the two temperature probe connectors are located directly above the 24-pin ATX and USB 2 headers.

So, what exactly stands out compared to Gigabyte’s previous X399 offerings? The Aorus Xtreme melds the flare of the Aorus 7 and Designare EX boards with the inclusion of the integrated X399 illuminated heat spreader and NVMe heatsink. I don’t prefer that approach, but the removal of the heatsink was straightforward and the provided aesthetic is a definite improvement for the Aorus brand. A simple workaround: Just don’t populate the bottom slot for the best of both worlds!

Again, taking a page out of the Designare EX’s playbook, the Aorus Xtreme deploys a matte black backplate for some additional swagger for custom builders or benchtop systems. In our Corsair 570X case, we found a little bit of resistance when installing the board, but otherwise it works as intended. Also, let’s not forget about that voltage regulation (VReg) heatsink: It’s super effective and will get more attention below in the overclocking section.

So, on the surface, the Gigabyte X399 Aorus Xtreme checks all the boxes for a flagship motherboard on a flagship chipset. Sure, it might not have an extra-gaudy outer shell or over-marketed VReg circuits, but it does a good job of mixing what we loved about the Aorus Gaming 7’s illumination and the Designare EX’s silver aesthetic, to make a beautiful X399 board. Before we get to the benchmarks, let's re-visit the UEFI.

MORE: Best Motherboards

MORE: All Motherboard Content

UEFI and Test Configuration

UEFI – Same ‘ol X399 Adds Support For 2000 Series

If we were to rank overall UEFI experience against the main motherboard competitors, Gigabyte would land in a startling fourth place with its lackluster looks and at times difficult-to-navigate menus (at least on AMD versions). However, ithe UEFI here is functional and provides a consistency that some vendors do tend to lack across their motherboard stack. For our previous analysis of Gigabyte’s X399 UEFIs, feel free to dig into some of our previous articles.

With the introduction of the Threadripper 2000 series processors, newer versions of AGESA are needed to support this silicon. And with our testing we used the F4d version (which has now been replaced with F5c and AGESA 1.1.0.2), so your mileage may vary. Largely, the main M.I.T. menus remain unchanged and most of our previously used features still live in the same respective subtrees as our prior reviews. The Advanced Frequency Settings menu hosts most of the multipliers for both CPU and memory, and Advanced Memory Settings opens up the whole menu of timing options for DRAM.

But what makes this board Xtreme? Given the emphasis on power delivery and cooling, we went ahead and dialed up some of our favorite tweakable settings to their limits to see how far the UEFI would let you go. CPU clock speeds are settable to an impressive 6.375 GHz on the top end and DDR4-4400MHz is the fastest canned frequency for DIMM sticks. Along with base multipliers, impressive voltage settings and protections are available on this board, with programmable loadline, separate voltage and current protection settings, and phase control settings for switching rates and modes for how the regulators react to dynamic loads.

For the users who don’t want to overclock the old-school, Precision Boost Overdrive (PBO) is enabled on 2000 series Threadripper chips, and this motherboard boasts some impressive limits as well. For more information about what these values mean, look to our overclocking section below or the motherboard section of our 2990WX review. Base settings for PBO can be adjusted from the M.I.T. Advanced Frequency Settings menu and can be either Auto, Disabled, Enabled, or Manual.

The Manual menu enables us to bypass the board’s default PPT, TDC, and EDC values for power, temperature, and current limits respectively. Additional PBO settings can be found in the Peripherals menu, where the PBO Scalar can be set, and we found that the maximum value we could set was 5,000. Lastly, jumping back from the PBO configuration, we can adjust the TDP limit in our UEFI to a maximum of 10000W and a temperature limit of 1000C. Granted, I don't understand where these limits apply in AMD’s grand flow-chart of power management, but know that those knobs are there for tinkering.

Test System Configuration

Sound	Integrated HD audio
Network	Integrated gigabit networking
Graphics Driver	GeForce 399.24

Comparison Products

ASRock X399 Professional Gaming

ASRock X399M Taichi

Gigabyte X399 Aorus Gaming 7

Benchmark Settings

Synthetic Benchmarks & Settings
PCMark 8	Version 2.7.613Home, Work, Storage, Applications (Adobe & Microsoft)
3DMark 13	Version 4.47.597.0Test Set 1: Skydiver, 1920x1080, Default PresetTest Set 2: Firestrike, 1920x1080, Default PresetTest Set 3: Firestrike Extreme, 2560x1440 Default Preset
SiSoftware Sandra	Version 2016.03.22.21CPU Arithmetic, Multimedia, Cryptography, Memory Bandwidth
DiskSPD	4K Random Read, 4K Random Write128K Sequential Read, 128K Sequential Write
Cinebench R15	Build RC83328DEMOOpenGL Benchmark
CompuBench	Version 1.5.8Face Detection, Optical Flow, Ocean Surface, Ray Tracing
Application Tests & Settings
LAME MP3	Version 3.98.3Mixed 271MB WAV to mp3: Command: -b 160 --nores (160Kb/s)
HandBrake CLI	Version: 0.9.9Sintel Open Movie Project: 4.19GB 4K mkv to x265 mp4
Blender	Version 2.68aBMW 27 CPU Render Benchmark, BMW 27 GPU Render Benchmark
7-Zip	Version 16.02THG-Workload (7.6GB) to .7z, command line switches "a -t7z -r -m0=LZMA2 -mx=9"
Adobe After Effects CC	Release 2015.3.0, Version 13.8.0.144PCMark-driven routine
Adobe Photoshop CC	Release 2015.5.0, 20160603.r.88 x64PCMark-driven routine (light and heavy)
Adobe InDesign CC	Release 2015.4, Build 11.4.0.90 x64PCMark-driven routine
Adobe Illustrator	Release 2015.3.0, Version 20.0.0 (64-bit)PCMark-driven routine
Game Tests & Settings
*Ashes of the Singularity: Escalation*	Version 2.50.28527High Preset - 1920x1080, Mid Shadow Quality, 1x MSAACrazy Preset - 1920x1080, High Shadow Quality, 2x MSAAHigh Preset - 3840x2160, Mid Shadow Quality, 1x MSAACrazy Preset -3840x2160, High Shadow Quality, 2x MSAA
*F1 2015*	2015 Season, Abu Dhabi track, Rain1920x1080 - UltraHigh Preset, 16x AF3840x2160 - UltraHigh Preset, 16x AF
*Metro: Last Light Redux*	Version 3.00 x64High Quality, 1920x1080, High TesselationVery High Quality, 1920x1080, Very High TesselationHigh Quality, 3840x2160, High TesselationVery High Quality, 3840x2160, Very High Tesselation
*The Talos Principle*	Version 2672521920x1080 - Medium Preset, High Quality, High Tesselation, 4x AF1920x1080 - Ultra Preset, VeryHigh Quality, VeryHigh Tesselation, 16x AF3840x2160 - Medium Preset, High Quality, High Tesselation, 4x AF3840x2160 - Ultra Preset, VeryHigh Quality, VeryHigh Tesselation, 16x AF

Up today for comparison are only two brands, Gigabyte and ASRock. But these boards have a purpose in our charts. The Aorus Xtreme is obviously what we’re interested in, but we need the X399M and X399 Professional Gaming from ASRock for their 2990WX data on our recently updated testbed. Also, we are throwing the Gigabyte X399 Aorus Gaming 7 into the ring due to how pleased we were with its performance. Another angle for analysis: How do the top-tier boards we’ve reviewed (ASRock Pro Gaming and Gigabyte Xtreme) compare? There's a lot to unpack, so let’s get started.

MORE: Best Motherboards

MORE: All Motherboard Content

Benchmark, Power and Thermal Results

1950X: Synthetics, Applications, and Games

PCMark has its own issues with large core count processors that are remedied with PCMark 10. But for our purposes, the Gigabyte X399 Aorus Xtreme performs admirably with the Office workload, but struggles when compared to the X399M Taichi and Aorus Gaming 7 in the remaining workloads. The pure synthetic Sandra suite, however, begins to favor the flagship boards from both vendors, with minor wins across the suite until we get to the Memory Bandwidth test. This could be an indicator that these boards are routed with a bit more care regarding signal integrity, with nearly 4 percent higher scores when compared to the X399M and Gaming 7.

Our storage benchmark is a mixed bag of results across the vendors. But for the averages, the X399 Aorus Xtreme manages to place 2nd with a solid performance in random reads and writes. Everyone’s favorite benchmark, Cinebench R15, has become the de facto standard in terms of multi-core performance in the industry. And it’s a shame to see the Aorus Xtreme place “so poorly” against the competition. However, the spread is small and subsequent executions would likely move any of these boards up or down the result sheet. Regardless of synthetic render performance, CompuBench results for the X399 Aorus Xtreme are very positive, with nearly a two-percentage-point lead over the second-place board when all scores are averaged.

As our testbench evolves, Time Spy and Time Spy Extreme will better exercise a system like ours. But for our purposes today, Skydiver and Firestrikes will have to do. Skydiver’s 720p resolution is likely too-low of a load to justify on our system, and the Aorus Xtreme struggles to keep up in the Combined and Graphics workloads. However, Firestrike and Firestrike Extreme look better for the Xtreme, with only negligible losses but measurable leads in Combined Firestrike performance.

On the application front, results are as expected in the Handbrake, Blender and Lame64 workloads, which AMD touts as Threadripper’s strong suits. Adobe performs well on this instance of our test bench with the X399 Aorus Xtreme coming out roughly 1.5 percentage points ahead of the X399M Taichi. Lastly, 7-Zip also shows strong performance with the Aorus Xtreme. However, the gap is less than one percent between first and last place.

Without repeating ourselves about X399’s gaming pedigree, let’s just get the games out there. Ashes of the Singularity at 1080p and 4k perform well with the Gigabyte X399 Aorus Xtreme, placing middle of the pack in all resolutions and details.

F1 2015 is limited by the number of cores/threads that can be used. So by deploying only 28 cores in Creator Mode through Ryzen Master and F1’s XML interface the Aorus Xtreme slips to the bottom half of performers as compared with this sample set. There are notable wins for the Gigabyte X399 Aorus Gaming 7 and the ASRock X399M Taichi in this workload.

Metro Last Light continues to be a good general metric for gaming, and the 1950X Threadripper-equipped X399 systems show a tight range of average framerates at both 1080p and 4k. And unfortunately, both top-shelf motherboards today lag the higher-value boards by nearly six percentage points at both resolutions and detail settings.

While we've seen strange results from The Talos Principle in the past, this time around the title provided expected results similar to our Metro workload. In other words, on this test the flagship boards struggle, lagging behind by up to 10 frames per second (fps) at 1080p and the High preset, and that trend continues at Ultra and 4K settings.

2990WX: Synthetics, Applications, and Games

Again, we’re limited in hardware that that we've tested with the 2990WX, so we’ll only be looking at the Gigabyte X399 Aorus Xtreme, ASRock X399M Taichi, and ASRock X399 Professional Gaming boards today. For synthetics, these 32-cores chew through Sandra as expected, with the Aorus Xtreme coming out ahead in the Arithmetic but falling behind in Multimedia, Crypto, and Memory Bandwidth tests by less than one percentage point.

IOmeter again shows each board excelling in different metrics of this test, and the Aorus Xtreme places well with all except the 4K Random Read performance sub-test. As we saw with the 1950X data, results scale very well in Cinebench with the 32 cores, but the Gigabyte X399 Aorus Xtreme does slip by a half a percentage point when compared to the X399M Taichi. And in a similar vein as the 1950X data, CompuBench performs well with the Xtreme motherboard and this 2990WX processor.

When combining the Gigabyte X399 Aorus Xtreme with the 64-thread processor, Handbrake pulls a 1-minute lead compared to the X399 Professional Gaming, but other workloads show smaller leads and losses for the Gigabyte board. Still, the spreads are all within two percentage points of one another.

As for gaming synthetics, more cores unfortunately don’t translate to better performance. Skydiver and regular Firestrike show the X399 Aorus Xtreme doing well in Combined and Physics scores, respectively. Firestrike Extreme hits the Xtreme board hard, with sizable losses in Graphics and Physics while still it still performs well with the Combined metric. On a side note, the average Combined scores for Firestrike with the 2990WX are six percent below the 1950X data we previously discussed. You definitely don't want to opt for one of these high-core-count CPUs specifically for gaming.

All three boards today perform pretty similarly to each other in games with instances where a few frames translate into larger percentage point deltas from the average. Starting with Talos, 2fps at the reported average frame rate separate 4K results at both detail settings, and at 1080p we see the X399 Aorus Xtreme lead the pack at both details presets.

Metro sees the opposite effect, where the X399 Aorus Xtreme lags by nearly 10fps at 1080p and high preset and bounces back to expected results at elevated details and resolutions. We clearly see GPU bottlenecking at 4k, where the 1950X was able to perform at similar framerates at the highest detail settings.

F1 2015 favors the frequencies of our 1950X, but for the 2990WX the Aorus Xtreme board suffers a 4fps loss at 1080p and wins the 4K average frame rates by a less-than-impressive 0.5fps.

Ashes also appears to favor speed instead of cores in our benchmarks, and the Aorus Xtreme serves up better performance at the higher detail settings at 1080p. However, 4K shows less than stellar Max Batch performance which drags down its percentage metrics in this workload.

In short, this board works well out of the box with either the 1950X or 2990WX with nearly 100 percent performance reported in all metrics except gaming, when comparing the averages. From a raw performance perspective, gaming on the 1950X goes to the Gigabyte X399 Aorus 7 and the ASRock X399M Taichi makes a clean run, scoring above 100 percent performance on all metrics. However, loading in the more power-hungry processor shakes things up just enough to show the ASRock X399 Professional Gaming and Gigabyte X399 Aorus Xtreme taking wins in most cases, with the former favoring gaming and latter favoring applications respectively.

Power, Thermals & Efficiency With Both Processors

Both processors are binned differently from a TDP perspective, so drawing direct comparisons between the two can be difficult. However, with our ample sample size, it is easy to compare our samples against each other. With the 180W part, the Gigabyte X399 Aorus Xtreme draws roughly four percent more power on average than the ASRock X399M Taichi, and trails the other Gigabyte by 2.4 percent. With the 250W 2990WX, the Gigabyte manages to tighten the gap to three percent on average by drawing in 5-10 extra watts at each test condition. Regardless, the Aorus Xtreme does well for what its purpose is.

This is the paragraph that will set the Aorus Xtreme apart from the competition. Judging solely by the 1950X data, it would appear that the flagship Gigabyte board runs warmer in all regards, despite the beefy VReg cooler and this particular CPU cooling solution. However, the tables turn when we increase the load and engage the same workload. We see a remarkable 18.2C delta between the ASRock MOSFETs and the Aorus Xtreme, as reported by HWiNFO. If we factor in ambient temperatures, the ASRock boards are running nearly at the top of their safe temperature threshold at 110C! In comparison, Gigabyte’s solution provides us with almost 30C of thermal headroom on the MOSFET, providing an opportunity to overclock the processor without additional cooling or airflow improvements. This will pay off in in our overclocking and conclusion sections soon.

Onto efficiency. With stable power measurements and performance numbers all within a few percentage points, we can easily draw definitive conclusions this go-around. At lighter loads on the 1950X, the Gigabyte X399 Aorus Extreme suffers in efficiency due to the extra wattage pulled at all system load levels. As we populate more load-intensive CPUs (2000 series), this improves by 1.3 percent, but doesn’t beat the X399M or Professional Gaming boards. This is unfortunate for Gigabyte, but not a glaring defeat by any stretch.

MORE: Best Motherboards

MORE: All Motherboard Content

Overclocking, Value and Conclusion

Clearly, pushing Threadripper to its limit is the goal you’re looking at the Gigabyte X399 Aorus Xtreme. With a 1950X installed in the system, we hit similar overclocks to the other samples while maintaining similar processor temperatures. As we see with the temperature charts, our VRegs maintain safe and optimal conditions for operation, thus providing for zero PROC_HOT or thermal throttles of the regulators. In other words, there's plenty of regulator thermal threshold for builders with custom loops or exotic cooling.

Swapping in the 2990WX, we have several avenues of overclocking, so let’s start with traditional voltage and multipliers. Given the drastic voltage changes, we stick with auto voltages for the first part, and we’re already hitting 100C on the MOSFETs at 3300MHz. Any higher, and we start to hit PROC_HOT throttling. Disabling PROC_HOT, we achieved 3500MHz for short runs. But to be safe, we dialed it back a bit to not cook our $1800 processor. After getting comfortable with the auto-voltage settings, we dialed in a manual voltage of 1.2V for the processor and immediately hit processor temps in the 90C range at 3500MHz. Only by reverting to the “normal” voltage setting are we able to maintain decent operating conditions for runtimes in the sub-1 hour range. For extended use cases, we recommend keeping the overclocks a little slower unless running a custom loop with ample radiator capacity.

Memory overclocking is consistent across both processors, and we were able to achieve DDR4-3200MHz with little effort. In the instance any instability occurres, simply adjusting the loadline of the memory rail to a higher setting typically will grant stability for AIDA64 tests. To watch the entire overclocking event, you can head to my Twitch page to learn a bit more about overclocking the 2990WX processor to 4.1GHz.

In our ASRock X399M Taichi review, we commented on our overclocking results versus what other outlets are reporting and wanted to expand upon that idea a bit. For my overclocking tests, I tend to use Prime95 for worst-case power and voltage conditions to publish conservative values. Some outlets like AIDA64s FPU or CPU workloads, which come with the added benefit of pretty charts and simplistic benchmark tweaking. Others prefer Cinebench, as it gives impressive results while running quickly.

The above picture helps illustrate these different workloads and their impact on system temperatures. From left to right, the first plateau we see is the Prime95 run where we hit PROC_HOT throttling, which helps keep our regulators safe. The 2nd ramping curve (surrounded by white vertical-dashed lines) shows AIDA’s FPU workload and at the same voltage and frequency. Our temperatures don’t induce thermal throttling, which might suggest higher multipliers are obtainable. Lastly, the downward-jaggy line is our custom Cinebench looping routine that shows the workload’s rapid completion. This test gives the Gigabyte X399 Aorus Xtreme’s VReg cooling solution time to do its job while maintaining a stable (and much cooler) processor temperature.

With that out of the way, let’s shoot for the moon with our Cinebench looper workload. Running at 3600MHz and utilizing the UEFIs voltage bias adjustments (set at “normal” mode) we easily pulled off roughly 5300 points. Cautiously ramping up the multiplier and voltage bias to 250mV, we can reach 4.0 GHz and a resulting score of 5820 points at a toasty 87C on Tdie in the latter loops. Squeaking in a non-looper run, we can hit 4.1GHz at these same conditions, though these are not safe for everyday use.

Let’s not forget the 2000 series processors’ Precision Boost Overdrive feature. Setting to Enable and Auto, this 2990WX boosts the multipliers up to 42, but can’t run all of them at the same time, resulting in a high score of 5650 and trailing down to 5300 points. Setting the PBO to Manual with the 2990WX results in a max score of 5400 that trails down to 5250 points, suggesting that the auto setting is higher. Bumping up the setting to 5000 results in mixed results that we will need to dig into more as we learn more about the platform.

Value & Conclusions

As expected from a top-tier flagship product, this board will never win when it comes down dollars. Even on sale, both the Gigabyte X399 Aorus Xtreme and ASRock X399 Professional Gaming fall short on “value,” and need to be looked at through different lenses. Since both boards are comparable from a connectivity and options perspective, we have to look a little deeper for recommendations.

For an out-of-the-box pro-sumer who needs 10Gb connectivity, isn't that concerned about performance, and plans to use either 1000 series or lower-wattage 2000 series Threadrippers, either board would work, although the ASRock Pro Gaming would appear to be the value choice there. Power users, on the other hand, will want to pick up the Gigabyte X399 Aorus Xtreme, because of its impressive capability of this board to manage higher-wattage workloads and parts when combined with custom loops and ample radiator space.The VReg cooler alone eliminates the need for a monoblock, and could tip the high-wattage value proposition back in Gigabyte’s favor for a fully equipped motherboard.

If 10Gb is not needed, we still recommend the Gigabyte X399 Aorus Gaming 7 for full-sized builds. And the ASRock X399M Taichi still opens the door for HEDT builders on the value front. Still, if overclocking high-end processors is your thing, it’s hard to look past the heatsinks on the Aorus Xtreme!

MORE: Best Motherboards

MORE: All Motherboard Contentd

ASRock X399M Taichi Review: TR4 Goes Tiny

Jacob Terkelsen — Sat, 01 Dec 2018 14:00:00 +0000

Layout and Features

Despite its surprising combination of a compact Micro-ATX form factor and well-established powerhouse X399 chipset, the ASRock X399M (currently $299 / £233 on Newegg, with an extra $20 / £30 rebate to sweeten the deal) delivers spot-on performance in line with larger boards. And there’s plenty of potential for overclocking if that’s your thing. However, those looking for the monster Threadripper build experience with loads of graphics cards and high-speed storage should look elsewhere.

We’ve tested the other X399 motherboard offerings from ASRock, including full-size Taichi and Fatal1ty models, in previous reviews and have seen comparable performance, overclocking, and feature sets across the product stack. The last ASRock Threadripper board to land on our test bench was the X399M Taichi. So when we received the box for the X399M we were at first perplexed about the board's size.

On paper, X399 screams for extra memory modules, lots of PCIe cards, and loads of storage connectors to make a monster of a workstation or pro-sumer production rig. So why would we intentionally pick a product that sacrifices many of those critical components and slots to fit in a small case? In order to cram the cores of both the 1000- and 2000-generation Threadripper CPUs into a smaller form factor, IO has to be sacrificed, without neglecting power delivery and cooling capability for AMD's powerful-power-hungry Threadripper CPUs. But ASRock still manages to squeeze quite a bit into the 9.6 x 9.6-inches of board space here. First, let's take a look at the specs.

Specifications

Socket	TR4
Chipset	AMD X399
Form Factor	Micro ATX
Voltage Regulator	8+3 Phases
Video Ports	✗
USB Ports	10Gbps:(1) Type A, (1) Type C5Gb/s: (8) Type A
Network Jacks	(2) Gigabit Ethernet
Audio Jacks	(5) Analog, (1) Digital Out
Legacy Ports/Jacks	(1) PS/2
Other Ports/Jack	(2) SMA Antenna, (1) BIOS Flashback
PCIe x16	(3) v3.0 (x16/x16/x16)
PCIe x8	✗
PCIe x4	✗
PCIe x1	✗
CrossFire/SLI	4x / 4x
DIMM slots	(4) DDR4
M.2 slots	(1) PCIe v3 x4 / SATA3 42/60/80/110mm(1) PCIe v3 x4 / SATA3 42/60/80mm(1) PCIe v3 x4 / SATA3 30/42/60/80mm
U.2 Ports	(1) PCIe v3 x4, SFF-8639 (disable M2_1)
SATA Ports	(8) 6Gb/s
USB Headers	(2) 5Gb/s Type-C, (2) USB 2.0
Fan Headers	(5) 4-Pin
Legacy Interfaces	✗
Other Interfaces	(1) RGB-LED, (2) Front panel Audio, (1) TPM
Diagnostics Panel	Numeric
Internal Button/Switch	CMOS Clear, Power, Reset, Xtreme OC
SATA Controllers	Integrated (0/1/5/10)
Ethernet Controllers	(2) Intel® I211AT
Wi-Fi / Bluetooth	Intel® 802.11ac WiFi Module, Bluetooth 4.2 / 3.0
USB Controllers	✗
HD Audio Codec	ALC1220
DDL/DTS Connect	✗ / Yes
Warranty	3 Years

ASRock's design acumen is in overdrive on this board, no-doubt aided by the company's experience building Mini-ITX boards for the more mainstream / consumer market. By sacrificing half of the DIMM slots (which still leaves you with four), the socket can be shifted towards the back panel to squeeze in the M.2 drives that are critical for high-performance storage drives and arrays. Also, PCIe slots are either removed or squeezed on the bottom half of the board, which limits dual-slot graphics card deployments to just two. On the surface, this appears to be a no-go for serious X399 builders. But for an average builder, populating four DIMM slots and two graphics card slots is probably pushing the limits of a reasonable build budget--especially given the high price of RAM and Nvidia's 20-series GPUs. Plus, SLI / Crossfire builds are supported less-and-less these days, both by developers and Nvidia / AMD. So maybe we could look at the ASRock X399M Taichi as the perfect entry-level X399 board?

The X399M sports the same voltage regulator (Vreg) components as the Taichi and Fatal1ty boards, while shifting the alignment of the heat spreader to accommodate the ATX mounting screws and reduced distance of the bend around the DIMM slots. This is a great feature from a raw power delivery perspective, but the same Vreg cooling issues we’ve seen on previous boards are carried over as well. We'll revisit power delivery in our overclocking section, but for the most part, this design works.

With eight USB 3.1 5Gbps (Gen1) ports, two USB 3.1 10Gbps (Gen2, one Type-A, one Type-B) ports, two gigabit Ethernet jacks, five analog and one digital audio port, a PS/2 port and connections for SMA Wi-Fi antennas, the lack of sacrifices for X399M back-panel implementation should be a relief to content creators or other accessory-laden users. Despite the reduced planar size, two USB 5Gbps, two USB2.0, one RGB, and five 4-pin fan headers are available on the board which, despite its smaller size, favors full-sized case configurations. This Micro-ATX board is starting to sound more and more enticing.

Other notable inclusions for the X399M Taichi are the Purity Sound 4 (essentially premium audio circuitry, capacitors and shielding) and the ALC1220 codec (just like the full-size Taichi), built-in Intel AC-Wireless, and overclocking-friendly features like power / reset buttons and an Xtreme OC switch.

As we mention above, the X399 standard inclusion of three M.2 slots are available here, but laid out a bit differently than the larger X399M’s sibling boards. Also, U.2 remains on the board with similar stipulations, such that connecting U.2 disables the M2_1 socket. Eight SATA 6Gbps ports are available through right-angled brackets, though clearance with smaller chassis could be problematic.

To pull a few universal complaints most every Mini-ITX review: Cramped spaces are certain to be a sticking point for some builders. Important fan and EPS connectors are tightly crammed together, which makes installation into smaller chassis even more of a challenge. The Dr. Debug LED is shifted from the bottom-right quadrant to above the 24-pin ATX header, which might be hard to see in some board orientations. Power and reset buttons are available, but they're tiny. Lastly, the reduced real estate leaves the board looking slightly boring by “gamer” standards, but a subtle glow below the chipset heatsink does provide the user with customizable coloration.

Typically, we cover the UEFI of a product in its own page in an article, but little has changed on the ASRock X399 front here. Those looking for BIOS details should check out the Taichi and Fatal1ty UEFI pages for a close look at the menus and options.

MORE: Best Motherboards

MORE: All Motherboard Content

Test Configuration and Article Specifics

For the most part, our X399 test bench programs haven’t changed over the past year. We still have instances where applications are broken due to software incompatibilities with high core counts. We are, though, happy see that the introduction of even more cores with our second-generations Threadripper 2990WX doesn’t break more test programs.

Test System Configuration

Sound	Integrated HD audio
Network	Integrated gigabit networking
Graphics Driver	GeForce 399.24

Comparison Products

ASRock X399 Taichi

Gigabyte X399 Designare EX

MSI X399 Gaming Pro Carbon AC

Benchmark Settings

Synthetic Benchmarks & Settings
PCMark 8	Version 2.7.613Home, Work, Storage, Applications (Adobe & Microsoft)
3DMark 13	Version 4.47.597.0Test Set 1: Skydiver, 1920x1080, Default PresetTest Set 2: Firestrike, 1920x1080, Default PresetTest Set 3: Firestrike Extreme, 2560x1440 Default Preset
SiSoftware Sandra	Version 2016.03.22.21CPU Arithmetic, Multimedia, Cryptography, Memory Bandwidth
DiskSPD	4K Random Read, 4K Random Write128K Sequential Read, 128K Sequential Write
Cinebench R15	Build RC83328DEMOOpenGL Benchmark
CompuBench	Version 1.5.8Face Detection, Optical Flow, Ocean Surface, Ray Tracing
Application Tests & Settings
LAME MP3	Version 3.98.3Mixed 271MB WAV to mp3: Command: -b 160 --nores (160Kb/s)
HandBrake CLI	Version: 0.9.9Sintel Open Movie Project: 4.19GB 4K mkv to x265 mp4
Blender	Version 2.68aBMW 27 CPU Render Benchmark, BMW 27 GPU Render Benchmark
7-Zip	Version 16.02THG-Workload (7.6GB) to .7z, command line switches "a -t7z -r -m0=LZMA2 -mx=9"
Adobe After Effects CC	Release 2015.3.0, Version 13.8.0.144PCMark-driven routine
Adobe Photoshop CC	Release 2015.5.0, 20160603.r.88 x64PCMark-driven routine (light and heavy)
Adobe InDesign CC	Release 2015.4, Build 11.4.0.90 x64PCMark-driven routine
Adobe Illustrator	Release 2015.3.0, Version 20.0.0 (64-bit)PCMark-driven routine
Game Tests & Settings
*Ashes of the Singularity: Escalation*	Version 2.50.28527High Preset - 1920x1080, Mid Shadow Quality, 1x MSAACrazy Preset - 1920x1080, High Shadow Quality, 2x MSAAHigh Preset - 3840x2160, Mid Shadow Quality, 1x MSAACrazy Preset -3840x2160, High Shadow Quality, 2x MSAA
*F1 2015*	2015 Season, Abu Dhabi track, Rain1920x1080 - UltraHigh Preset, 16x AF3840x2160 - UltraHigh Preset, 16x AF
*Metro: Last Light Redux*	Version 3.00 x64High Quality, 1920x1080, High TesselationVery High Quality, 1920x1080, Very High TesselationHigh Quality, 3840x2160, High TesselationVery High Quality, 3840x2160, Very High Tesselation
*The Talos Principle*	Version 2672521920x1080 - Medium Preset, High Quality, High Tesselation, 4x AF1920x1080 - Ultra Preset, VeryHigh Quality, VeryHigh Tesselation, 16x AF3840x2160 - Medium Preset, High Quality, High Tesselation, 4x AF3840x2160 - Ultra Preset, VeryHigh Quality, VeryHigh Tesselation, 16x AF

With the introduction of the 32-core 2990WX and other 2000 series Threadrippers, existing X399 motherboards technically support the power limitations of these latest processors with limited boosting abilities. To test that specification against this tiny board’s stature, we've introduced the 2990WX into our testing regimen and will also be featuring it and this board in an upcoming Micro-ATX Monster build feature. Our goal today though is to see how this motherboard handles our standard test suite as well as this worst-case power scenario with AMD's latest 32-core monster chip.

Our pristine tempered-glass Corsair 570X case has fallen prey to the dangerous combination of pre-adolescent hands and permanent markers. It still works, so leave us tips on how to remove markers in the comments! On a more serious note, the Antec 1200W is a necessary inclusion to our test rig, since overclocking a 2990WX can bring system load numbers into the 800W range. Tacking on the GPU might even bring even this power supply to its knees. Our Flare X kit still provides DDR4-3200 speeds and overclocks fairly well in the right system. System storage is satisfied with the Toshiba RD400 256GB NVMe drive. Cooling our processor today is the Fractal Design S36 Celcius, so we’ll be testing its capability of cooling these processor configurations. We’re still using the GTX 1080 card because RTX cards are still hard to come by--as are games that currently support ray tracing.

MORE: Best Motherboards

MORE: All Motherboard Content

Synthetics, Applications and Gaming Performance

1950X Data: Synthetics, Applications & Games

Our ASRock X399M and Gigabyte Designaire boards appear to enjoy favorable wins with the Home, Work, and Office benchmarks, likely in part due to driver or chipset changes over the past year, while Adobe and Storage benchmarks appear to be unaffected for the most part. However, Sandra shows consistent results in the Arithmetic, Multimedia, and Memory suites regardless of test execution date. Cryptography does show slight advantages to the more recent test samples, though the 5 percent spread is nearly negligible.

Everyone’s favorite multithreaded synthetic test these days is Cinebench. And the tiny X399M Taichi enjoys a second place win in single-threaded performance there, but slumps to third place with multi-threaded performance. Compubench shows similar preference in the Video Process workloads. But beyond that, all systems operate within range of 4 percent from first to the last place. IOmeter is our preferred storage benchmark for the Toshiba RD400 drives and there the X399M sees negative marks across the board by upwards of 8 percent from the leaders. This begs the question of whether we should believe PCMark or IOmeter when it comes to storage speed.

We favor the Firestrike and Firestrike Extreme results from 3Dmark for a workstation platform, and the X399M Taichi comes out on top with the 1440p workload across all metrics. The original X399 Taichi also performs well at the 1080p setting in our testing, though its smaller brother starts to slip up with the lighter graphical workloads. Focusing on the CPU scores, all boards show a performance range of 3-5 percent across all workloads.

Our Application suite is by no means the sweet spot for Threadripper, but many of the jobs we run do favor the core counts at their disposal. Our CPU Blender workload shows a favorable advantage to the X399M board but lags by a half a second in the LAME encoding scenario. Adobe shaves 5 to 9 seconds off the After Effects workload and lags by fractions of seconds in Illustrator. The X399M is tied in the middle of the pack with 7-zip, which still favors clock rates rather than cores.

Our previous reviews have proven that gaming at mainstream resolutions is not the ideal use case for Threadripper, but the 1950X processor is still potent there. Our resident CPU-bound title, Ashes of the Singulatiry, shows very consistent results across the test samples, with the larger ATX Taichi taking some wins at the 1080p Crazy setting, but the MSI Pro Carbon and Designaire takes wins when we step up to 4K High / Crazy presets. The X399M Taichi takes back-of-the pack scores here, though this should not be discouraging considering the small 3.5 percent performance gap.

F1 2015 has troubles with core / thread counts larger than 30, so our 1950x doesn’t fully leverage its potential in this game--and that's even more more true with the 2990WX. At 1080p, both ASRock boards fall behind the MSI and Gigabyte contenders, but the story changes as graphical resolution quadruples. Again, each sample here is within a single-digit percentage range of its competitors.

The Talos Principle repeats our observations from previous test games, and the little Micro-ATX monster lands in either second or third place in a majority of test conditions. Framerates are still suitable for both high-refresh-rate monitors and larger-resolution screens. So X399 users can take breaks from editing video other work tasks by playing this puzzle game.

2990WX Data: Synthetics, Applications & Games

Shifting gears to the new big-boy CPU, the 2990WX sacrifices core frequency and power for those tasty threads, nearly doubling the capacity of our 1950X sample. Granted, it’s not a direct apples-to-apples comparison, but these tests should also give is an idea of how the 2950X could operate on this same platform.

At first glance, the 2990WX synthetic data looks very compelling on both the X399M and Professional Gaming boards, showing gains in all the PCMark workloads except Storage, Adobe CC, and Microsoft Office scenarios. The WX processor really struts its stuff with Sandra, showing excellent scaling in both the Arithmetic and Multimedia workloads by nearly 90 percent compared to the 1950X. Piling on the gains, Cinebench steams ahead with a 67 percent performance gain thanks to the core counts and the occasional boosts from Threadripper power management. However, those gains quickly erode with a 42 percent memory bandwidth degrade and as much as 50 percent lower crypto results. Despite the architectural differences between these processors, the X399M manages to stay ahead of the Fatal1ty Professional Gaming board in most instances.

Application data shows similar behavior to the Crypto and Memory bandwidth, where many of our applications just don’t care that we have an abundance of cores. The shining stars though are PCMark’s Adobe After Effects workload and our Blender CPU benchmark. Otherwise, more frequency helps the remaining tests.

We've said it before and we'll no-doubt say it again: You shouldn't be spending $1,800 USD on a 2990WX just to play games on it. Don’t get us wrong, gaming is still viable and highly enjoyable on the platform. But not many of the games in our suite (or current existence) utilize all of the extra cores, and are effectively bottlenecking our GTX 1080. Ashes generates around 40 percent less performance at 1080p here, and we do observe performance tighten a bit as the graphical load increases.

F1 at 1080p also sees roughly 30 percent lower frame rates at the Ultra-High preset, but 4K leans more heavily on the GPU and the delta shrinks to 8 percent. But, the X399M struggles compared to the Fatal1ty board at 4K.

Metro shares the same 1080p fate as the other games in our suite, where framerates lag between 25 and 35 percent, though the Fatal1ty does appear to manage higher average framerates (we do not factor in mins or maxes in this test). The only saving grace for the 2990WX in this suite is the 4K Very High detail setting, where all results are gated by the GTX 1080’s horsepower.

So, as hyped as we are that this X399M can drive and boost this processor accordingly, the 2990WX just isn’t well suited for our motherboard test suite. Even with the monstrous synthetic wins and modest application wins, gaming performance drags this build down.

MORE: Best Motherboards

MORE: All Motherboard Content

Overclocking: Two Stories, Similar Results

With successful 1950X benchmark results and compelling thermals and power numbers, we have an expectation that the ASRock X399M Taichi can hit similar CPU and memory frequencies to its larger ASRock siblings. We were able to dial in similar settings to our Taichi and Fatal1ty reviews by disabling PROC_HOT, voltage safeguards, lowering the loadline and landing at a VCore regulator setpoint of 1.2625V. We still observe instances of thermal throttling where the multiplier drops to 5.5 to keep VReg and CPU temperatures in check. Regardless, our Prime95 load during our overclocking is extremely stressful for these CPUs, thus keeping us from hitting that precious 4GHz overclock.

Memory frequency was a surprising defeat. We were only able to achieve AIDA64 stability with the DDR4-3066 setting through the UEFI with the 1950X. We typically see memory overclocks improve when slots are removed and brought closer together, translating into better signal integrity to the DIMMs. Another likely cause could be that some of our DIMMs don’t like to be in specific slots, proving that achieving higher clocks can be obtained through patience and perseverance instead of brute force.

Swapping in the 2990WX, we are already thermally limited by the VReg cooling solution, but we charged ahead and pushed the system further. Voltage setpoints are drastically different between the processors, so disabling PROC_HOT, MOSFET protections, and setting VCore to 1.05V led us to an extended Prime95 overclock of only 3225MHz. The above picture shows that 3450MHz can be set in the UEFI, but the VReg temps will go beyond the safe point for MOSFETs, making additional cooling necessary. On the surface, this seems like a defeat, but let’s take a step back.

Most media outlets are touting 4.0GHz or higher overclocks on their 2990WX samples. Upon further inspection, they are either running larger coolers or lighter workloads. As many outlets are posting Cinebench scores, we will write up a short 10-loop Cinebench multithreaded execution test that gives us several data points. First, we get that impressive “1 run” overclock that most outlets bump chests about. The second data point we gain is repeatability. A one-run overclock is only useful for that one run. As components heat up, voltages fluctuate, and different platform choke points are stressed with successive runs. On a related note, we want a workload that is going to be analogous when we start to engage Precision Boost Overdrive, which we will cover shortly.

Being ambitious, bumping the voltage up to 1.2V and setting the multiplier to 39.25 easily hits 5600 points with Cinebench across all loops and our regulator is averaging out at 74C, or 50C above ambient, which is well below our Prime95 setting. With that thermal headroom, we can bump the voltage up to 1.25V and crank our multiplier to 40.00 to achieve the elusive 4GHz point. Any higher on the frequency and we fail out on repeated loops, though we were able to hit 4.05 with a few successful loops before our system required more voltage for stability. Resulting scores are 5850 at 4GHz and pushing the 5900’s for the less stable frequencies.

If that all flew over your head, there is an easy way to hit 4GHz-like performance, and that is with Precision Boost Overdrive. NOTE: enabling this feature will void the warranty on your precious processor. Throwing risk into the wind, we void ASRock’s warranty and sure enough, setting PBO to AUTO gives us a Cinebench Looper score of ~5400 points for a 6% increase over stock settings. Not bad. Setting PBO to “ENABLE” engages ASRock’s aggressive board settings and our 1 run PBO score is an impressive 6016! However, successive loops prove the point that this system’s power and the thermal solution can’t maintain the load for longer durations. So our 10 loop average decreases to 5666 points. Though we don’t have the graph, it’s important to note that upon each loop, we decreased logarithmically to around 5200 points, which is very similar to our standard 2990WX stock result.

In short, the X399M overclocks as well as its larger brothers and shares the same flaws, mainly its VReg heatsink. In order to push the system harder, you'd need to employ a monoblock, which we'll be doing shortly in a build feature based around this board and the 2990WX. Cooling solution aside, this board can deliver the power required to push either Threadripper CPU well past standard operating conditions, but in doing so we see our Kill-A-Watt break the 800W threshold! Oh, and the 2990WX also hits DDR4-3200 memory speeds.

MORE: Best Motherboards

MORE: All Motherboard Content

Power, Thermals, Efficiency, Value and Conclusion

With the 1950X populated in the system, the ASRock X399M manages excellent idle and GPU torture metrics but gets docked for full CPU and average system load measurements by 6.4% when compared to the full-sized Taichi. Temperatures with this processor are also comparable to the intake radiator measurements of the Gigabyte and ASRock boards but the VReg cooler is toasty running 25 degrees Celsius hotter than its big brother. This spells concern as we install the higher wattage 32-core SKU.

The 2990WX is specified as a 250W part and loading the module with Prime95 Small FFTs pushes a total system power draw of 404W or 50% more power than the 1950X. It also feels very fulfilling to be running a processor at a higher power draw than our GPU for once, granted the GPU stress is burning more system power than the 1950X likely due to the larger processor. These larger percentages transfer to the combined torture test and the overall average metric, thus further proving that an 800W supply or better is required for full system stress. If my wife is reading this, sorry our power bill is higher this month.

Taking another glance at thermals, the 2000 series processors will always boost up to the thermal limit of 67.8C before throttling back the boost frequency of the processor. With that in mind, both ASRock boards boosted to similar levels before dialing back and equalizing around 3100MHz boost clocks. The downside here is the regulator temperatures, running 87C above ambient most certainly proves we’re pushing this board's thermal limits in its current configuration. This will be something that needs addressing when we assemble our little monster.

Efficiency is a tough sell for this platform and surprisingly the 1950X data is fairly tight across all the samples. Unfortunately, the X399M encounters elevated full-system load which tips it into the red for efficiency, but it is still relatively close. With that 32-core processor, power consumption is its middle name and clearly gets defeated by the 1950X considering the few wins in our benchmarks and drawing extra power from the wall during load. As for the X399M Taichi, it performs marginally better than the Fatal1ty board with the 32-core processor thanks to the slight power and performance advantage.

Final Thoughts

The ASRock X399M Taichi performs just as well as its larger ASRock sibling boards, and sacrifices connectivity, expandability, and some aesthetics to cram X399 into the Micro-ATX form factor. Putting aside the purist in me, this board makes sense from a value perspective. If all you're looking for is cores and speed, the added benefit of PCIe, DIMMs, and arguably storage are well worth the exclusion and cheaper price tag. At MSRP, this board squeaks by the larger Taichi by only $10 USD which is insignificant given the price of the rest of the hardware.

However, taking sales prices at the time of this publication, a savings of $20 USD could be the difference in a cooler selection, additional NVMe capacity, or even bumping up GPU tiers. Given our analysis of all the X399 boards from ASRock, the shared VReg cooling solution almost makes it essential to get a monoblock for extreme overclocking or prolonged high load use cases, so that $20 savings might be helpfully spent toward a custom cooling loop.

We put this tiny Taichi board through more than our usual spate of tests. And while a compact mATX Threadripper solution originally didn’t hold a whole lot of interest, the X399M Taichi opened our eyes to the possibilities of building a compact powerhouse PC. ASRock delivers full-sized performance in a smaller package at a price that is more palatable for a builder looking to dive into the HEDT scene. And to top it off, it can drive even the most demanding workloads and hardware that can be thrown at it, earning it a Tom’s Hardware Editor’s Choice award.

MORE: Best Motherboards

MORE: All Motherboard Content

ASRock X399 Fatal1ty Professional Gaming Motherboard Review: More Pro than Gaming

Jacob Terkelsen — Sun, 21 Oct 2018 13:00:00 +0000

Features & Layout

The ASRock Fatal1ty X399 Professional Gaming motherboard (priced around $390) performs just as well as the company's X399 Taichi, but has the added benefit of the 10Gb network controller. This board has some deficiencies from a gamer’s perspective, but would work well for a system that calls for 10Gb throughput and utilization of all the PCIe slots (something most boards can’t handle).

AMD successfully stole the tech communities’ attention at Computex 2018 with the announcement of the Threadripper 2000 series processors. And since the August release of the 32-core behemoth flagship chip, there has been plenty of positive press for the HEDT underdog. In our comparison against the Skylake-X class processors, AMD’s latest ecosystem only lags the Blue Team on gaming performance and motherboard selection, while generally costing a whole lot less per core and thread.

Specifications

Socket	TR4
Chipset	AMD X399
Form Factor	ATX
Voltage Regulator	8+3 Phases
Video Ports	✗
USB Ports	10Gbps:(1) Type A, (1) Type C5Gb/s: (8) Type A
Network Jacks	(1) 10Gb Ethernet(2) Gigabit Ethernet
Audio Jacks	(5) Analog, (1) Digital
Legacy Ports/Jacks	(1) PS/2
Other Ports/Jack	(2) SMA Antenna, (1) BIOS Flashback
PCIe x16	(4) v3.0 (x16/x8/x16/x8)
PCIe x8	✗
PCIe x4	✗
PCIe x1	(1) v2.0
CrossFire/SLI	4x / 4x
DIMM slots	(8) DDR4
M.2 slots	(2) PCIe v3 x4 / SATA3 42/60/80mm(1) PCIe v3 x4 / SATA3 30/42/60/80mm
U.2 Ports	(1) PCIe v3 x4, SFF-8639 (disable M2_1)
SATA Ports	(8) 6Gb/s
USB Headers	(2) 5Gbps(2) USB2.0
Fan Headers	(5) 4-pin
Legacy Interfaces	(1) COM
Other Interfaces	(2) RGB-LED, (2) Front panel Audio
Diagnostics Panel	Numeric
Internal Button/Switch	CMOS Clear, Power, Reset, Xtreme OC
SATA Controllers	Integrated (0/1/10)
Ethernet Controllers	(1) Aquantia® AQC107(2) Intel® I211AT
Wi-Fi / Bluetooth	Intel® 802.11ac WiFi ModuleBluetooth 4.2 / 3.0
USB Controllers	✗
HD Audio Codec	ALC1220
DDL/DTS Connect	✗ / ✗
Warranty	3 Years

Rather than cook up a new chipset like the Ryzen 2000 series launches, AMD chose to provide a better upgrade path for users in the field who need the extra cores or features of the 2000 series Threadrippers by keeping the existing X399 chipset and socket in place. For the most part, older vintages of X399 boards were overbuilt for the first-generation parts, but AMD officially supports only stock frequencies for its highest-end WX series processors with some older boards.

The ASRock Fatal1ty X399 Professional Gaming board shares several similarities to its Taichi brother. In fact, if the facade and aesthetics are removed, both the Taichi and Professional Gaming boards share nearly identical design and component placement across the board. Plastic IO shields, VReg heatsinks and header locations and styles are identical. In short, the only substantial difference between these boards is the inclusion of the Aquantia 10Gb network controller and the populated COM port header on the Professional Gaming board.

But enough about the Taichi; let’s take a close look at the Fatal1ty board. The ASRock Fatal1ty X399 Professional Gaming provides 10 USB ports through the back panel, which are wired out for eight USB3.0, one Type-A and one Type-C USB 10Gb/s. Five gold-plated analog audio jacks and a SPDIF port enable audio through the Realtek ALC1220 codec. And the board is bundled with the Sound Blaster Cinema 3 suite to leverage Creative’s software stack in conjunction with the ASRock’s audio circuitry and design.

Wireless network connectivity is granted using an Intel AC8265 2x2 802.11ac controller, and standard Ethernet is provided by two tried-and-true Intel I211AT gigabit controllers. The third RJ-45 connector is connected to the star of the show: the Aquantia AQC107 controller which enables 10GBASE-T connections (more on that later). Lastly, what would a Fatal1ty board be without support for a PS/2 port or the inclusion of BIOS Flashback on the back panel?

Moving clockwise across the planar, the top-left quadrant of the board houses the first of five four-pin PWM fan headers and a 4-pin EPS12V power connector. This board utilizes both the 4-pin and 8-pin to power its 8+3 phase regulator topology, with IR PWM controller and Dr. MOS integrated MOSFET drivers. Cooling this Vreg design is the same heatsinks we see on the X399 Taichi, which prove to be sufficient for the lower wattage Threadrippers: Additional airflow is needed when using CPU water blocks for an overclocked WX processor.

Quad-channel DDR4 is supported by default on Threadripper and this board provides support for ECC non-buffered DIMM modules. The right edge of the board houses the 24-pin ATX, 6-pin PCIe supplement power, and the first USB3.0 header. The 6-pin PCIe connector is required for multi-GPU configurations, and fortunately the angled connector helps keep cable management a little bit cleaner. Skipping storage ports for now, the bottom side contains the classic format of front panel headers, two USB 2.0, two 4-pin fans, front-panel audio, and COM port headers. For the overclockers among us, two 7-segment numeric debug LEDs provide Dr. Debug post codes, alongside power and reset buttons.

Storage on Threadripper comes as no surprise. Lots of PCIe lanes means more connections and options for system builders. ASRock opts to provide three M.2 and one U.2 connector for users who want to run both consumer grade M.2 and enterprise-class U.2 drives. However, if U.2 is enabled, the top M.2 is disabled, so don’t plan on migrating any 3-drive M.2 arrays and trying to leverage U.2 at the same time. Eight SATA ports enable standard SATA3 protocols with support for RAID 0/1/10.

Copying the Taichi board layout, the PCIe topology is routed out to four x16 connectors and one open-ended x1 connector. From top to bottom, they are wired as x16/x8/x1/x16/x8, where the larger ports are Gen3 PCIe and the x1 is Gen2. With the Aquantia chip providing 10Gb connections, this system really can leverage multi-GPU configurations without sacrificing Ethernet connectivity or imparting storage controller complications.

Given the “Gaming” moniker, this Fatal1ty product is a stark deviation from the Taichi’s aesthetic, favoring angles, gradients, and heavy use of darker grays and blacks. The X399 chipset is illuminated by LEDs emitting from behind the heatsink, and there are two RGB headers that support LED strips. We feel that the accompanying IO shields and regulator heatsinks make the Professional Gaming board more attractive than the Taichi model. But ultimately, both boards provide similar looks when fully populated.

All things considered, the ASRock Fatal1ty X399 Professional Gaming builds on the Taichi’s success with features that favor the workstation more-so than the gamer. For a gaming board, we would favor more fan/pump headers for water cooling and possibly removing the U.2 and Wifi card to reduce frequently unused features. For a workstation, relocating the power headers and spreading out some of the other components might improve serviceability.

MORE: Best Motherboards

MORE: All Motherboard Content

UEFI & Software

A lot of the functionality of the Fatal1ty board's UEFI is discussed in our previous ASRock X399 Taichi review, so we'll talk about what might be different from the older UEFI compared to more recent and beta versions. One feature I guarantee we didn’t talk about was BIOS Flashback. As testing commenced on this product, we are graced with the opportunity to test this specific feature. Sure enough, updating to a more recent version left our system in the dreaded 0% UEFI update state, which is always a panic-inducing situation. After consulting the documentation (page 30) we were able to resurrect this board with the latest UEFI. Luckily, the documentation is clear, so don’t be intimidated renaming files and formatting USB drives.

As with any Fatal1ty branded product, heavy use of reds and blacks fill the background of the UEFI and the more text-based interface. The Main, OC Tweaker, Advanced, HW Monitor, and other high-level menus operate just as before, and these similarities help us to dive deeper into the UEFI. An important note: This board was initially tested with UEFI version 3.1, which is no longer available on the product page, and newer versions (3.23B) implement performance enhancer features that we will not cover in this article.

The OC Tweaker’s CPU configuration menu has options for changing the BCLK, multiplier, and other frequency settings. Voltages, loadlines, and MOSFET protection options are also open to bypass, especially when attempting extreme overclocking efforts. One key feature for Threadripper 2000-series processors is the availability for Precision Boost Overdrive (PBO). By default, PBO is set to AUTO, which will only engage the standard performance boost features enabled by 2000 series processors. When set to Enable, ASRock implements its preferred settings for PPT (socket wattage), TDC (Vreg current draw), EDC (Vreg current capacity), and PTC (thermal limit). We recommend using Ryzen Master for monitoring system parameters while running PBO to compare runtime system usage against these predetermined values. Lastly, Manual mode enables the user to define the values to use for PBO, but we do not recommend these settings since ASRock's values are reasonable. Also note--and this is important--using PBO voids your warranty.

As with our ASRock X470 and B450 reviews, the Advanced menu now becomes critical for unleashing every ounce of performance for the newer processors. Core Performance Boost, NBIO, and CBS menus have options for various performance and power thresholds for the processor. We will talk more about the NBIO in the overclocking section, but know that adjusting the TDP values for the processor is accessible through this menu.

The Tool menu works similarly to the X399 Taichi's, allowing access to the RGB LED options. And we prefer the Breathing mode as presented through the UEFI. HW Monitor again functions like the other board, but note that the CPU temperature does appear to track with the Tctl option, so use caution when creating custom fan curves based on these values. These two menus highlight the love-hate relationship with ASRock and some reviewers. The interface is functional and provides access to the settings, though it can be quite intimidating and cryptic without the use of graphs and stylized prompts.

So, what makes this product a “gaming” board compared to other ASRock offerings? Bundled software this time around is one of the key differentiating factors and is primarily driven by the F-stream ecosystem. Once the packages are installed, gamers can opt to utilize the software to implement overlays, macros, and unlock more refresh rates for mice through the Fatal1ty mouse port. Though I myself don’t put stock in those features, serious gamers may find reason here to opt for the Fatal1ty board over the Taichi series.

MORE: Best Motherboards

MORE: All Motherboard Content

Test Configuration & 10Gb Teaser

Test Hardware

The Corsair Crystal Series 570X case still houses our Threadripper test bed, and it pairs well with the ASRock X399 Professional Gaming, as far as aesthetics are concerned. It's a looker, just like the board we're pairing it with.

Antec’s 1200W High Current power supply provides the juice to our Gigabyte GTX 1080 G1 Gaming GPU and Ryzen Threadripper 1950X and 2990WX CPUs. G.Skill provides the 32GB of DDR4-3200MHz RAM, and Toshiba's 256GB RD400 M.2 NVMe drive serves as our boot drive. Fractal Design’s Celsius S36 360mm water cooler is placed on the intake side of our chassis to provide cooling for our 190W and 250W processors.

Benchmark Settings

Synthetic Benchmarks and Settings
PCMark 8	Version 2.7.613Home, Work, Storage, Applications (Adobe & Microsoft)
SiSoftware Sandra	Version 2016.03.22.21CPU Arithmetic, Multimedia, CryptographyMemory Bandwidth
IOmeter	4k Random Read, 4k Random Write128k Sequential Read, 128k Sequential Write
Cinebench R15	Build RC83328DEMOOpenGL Benchmark
CompuBench	Version 1.5.8Face Detection, Optical Flow, Ocean Surface, Ray Tracing
3D Tests ansd Settings
3DMark 13	Version 4.47.597.0Test Set 1: Skydiver, 1920x1080, Default PresetTest Set 2: Firestrike, 1920x1080, Default PresetTest Set 3: Firestrike Extreme, 2560x1440 Default Preset
Application Tests and Settings
HandBrake CLI	Version: 0.9.9Sintel Open Movie Project4.19 GB 4k mkv to x265 mp4
LAME MP3	Version 3.98.3Mixed 271MB WAV to mp3Command: -b 160 --nores (160 Kb/s)
Adobe After Effects CC	Release 2015.3.0Version 13.8.0.144PCMark driven routine
Adobe Photoshop CC	Release 2015.5.020160603.r.88 x64PCMark driven routine (light and heavy)
Adobe InDesign CC	Release 2015.4Build 11.4.0.90 x64PCMark driven routine
Adobe Illustrator	Release 2015.3.0Version 20.0.0 (64-bit)PCMark driven routine
Blender	Version 2.68aBMW 27 CPU Render BenchmarkBMW 27 GPU Render Benchmark
7-Zip	Version 16.02THG-Workload (7.6 GB) to .7z, command line switches "a -t7z -r -m0=LZMA2 -mx=9"
Game Tests and Settings
Ashes of Singularity: Escalation	Version 2.50.28527High Preset - 1920x1080, Mid Shadow Quality, 1x MSAACrazy Preset - 1920x1080, High Shadow Quality, 2x MSAAHigh Preset - 3840x2160, Mid Shadow Quality, 1x MSAACrazy Preset -3840x2160, High Shadow Quality, 2x MSAA
F1 2015*	2015 Season, Abu Dhabi Track, Rain1920x1080 - UltraHigh Preset, 16x AF3840x2160 - UltraHigh Preset, 16x AF
Metro Last Light Redux	Version 3.00 x64High Quality, 1920x1080, High TesselationVery High Quality, 1920x1080, Very High TesselationHigh Quality, 3840x2160, High TesselationVery High Quality, 3840x2160, Very High Tesselation
The Talos Principle	Version 2672521920x1080 - Medium Preset, High Quality, High Tesselation, 4x AF1920x1080 - Ultra Preset, VeryHigh Quality, VeryHigh Tesselation, 16x AF3840x2160 - Medium Preset, High Quality, High Tesselation, 4x AF3840x2160 - Ultra Preset, VeryHigh Quality, VeryHigh Tesselation, 16x AF

With the inclusion of the Aquantia AC107 chip on this board, we must shoehorn a “test” beyond our standard motherboard scope. For our purposes today, we connected the 10GBASE-T interface into our workstation's BCM57810 dual-port 10Gb network card and configured them onto a 10.0.0.x private network. From there, we transfer a 14.5GB video file from a recent stream from one NVMe share point to the other computer and vice-versa.

Out of the box, the NIC was able to hit 986MB/s download speeds and provided smooth 1.10GB/s upload speeds to the other host. It is quite impressive that this NIC required no tweaking of settings such as jumbo frames or disabling certain flow parameters as mentioned in several guides online. Turn-key implementation of a high-speed interface is always a plus in our book.

Lastly, we managed to get our hands on a 2990WX processor for use with some of our X399 testing as we gear up for a future project, so we will try to keep that data isolated to our 32-core content section (page 5). All overall value, performance, and conclusions generated in this article are geared towards the standard 1000 and 2000 series processors.

MORE: Best Motherboards

MORE: All Motherboard Content

Synthetics, Applications & Games

Comparison Products

ASRock X399 Taichi

Gigabyte X399 AORUS Gaming 7

MSI X399 Gaming Pro Carbon AC

Synthetics and Applications

Typically, synthetics on this platform are fairly mundane. But right out of the gate, the ASRock X399 Professional Gaming surges ahead in PCMark, with impressive margins compared to all of the competitors. Sandra also plays well with the Fatal1ty board, with strong performance in all workloads. The Taichi also shows impressive results, indicating that the ASRock boards perform quite well for the prosumer crowd.

Looking at disk performance with IOmeter, our random reads and writes take a hit on this Toshiba RD400, which has roughly 75% of its rated endurance left. Mixed results are seen with Cinebench, where we observe average single-thread performance. But our multi-thread performance takes a subtle hit. CompuBench starts to shift more towards the GPU compute performance of the system and the ASRock X399 Professional Gaming trades wins with the MSI X399 Pro Carbon AC in various tasks. Honorable mention goes to the Gigabyte X399 Aorus Gaming 7, as it lands in second place in all but one instance in this workload.

With 3DMark, we see an interesting shift in trends where the ASRock Fatal1ty board starts to slip up with regards to combined and physics scores. Unfortunately, we can’t explain the Combined score drop in Firestrike, given the test results of the graphics and physics results being positive. Otherwise, a low-third place finish in 3DMark is underwhelming for the Professional Gaming board.

Applications, again, prove the strengths of the Threadripper platform, and the synthetic trends continue with solid wins for the Fatal1ty in Adobe, 7-zip, and some Blender runs. Although the scales in the graphs are misleading in some cases, each workload is within single-digit percentage points of each other, which shouldn’t impact overall performance metrics much.

Games and Performance

Remember how 3DMark started to show some deficiencies in the graphics department for the ASRock X399 Professional Gaming board? Those trends continue in all four of our different gaming runs, regardless of the resolution or graphical setting.

With Ashes of the Singularity: Escalation, the Gigabyte X399 Aorus Gaming 7 flies past the Fatal1ty board by up to 3 frames per second at 1080p, and the gap shrinks by a couple frames at UHD resolutions. The latter appears to be within sampling noise, so don’t draw too much concern. However, solid negative relative performance deltas will impact the Fatal1ty board's score moving forward.

As enjoyable as F1 2015 is at all resolutions and settings with our GTX 1080, we still see frame deficiencies within 5 percentage points of the Aorus board here. Shifting to 4K reduces the deltas by a couple frames here and there, but we still see a larger percentage point gap compared to the MSI X399 Gaming Pro Carbon AC.

For a change of pace, the ASRock X399 Taichi starts to take the lead with The Talos Principle at 1080 high, and the Gigabyte places first with both 4K presets. The ASRock X399 Fatal1ty board still suffers from minor frame deficiencies here, but keep in mind we are talking about fewer than nine frames per second.

Metro Last Light scoffs at us and decides to scatter its results across all test samples, with the ASRock Professional Gaming board taking a couple first- and second-place finishes. Another way to look at the Metro results is that comparing test methodology across several months of tests, the data reflects no preference to any particular software configuration or program update discrepancies. Our 4K results also reflect this observation, with Very High presets being identical across the field and only two frames separating the Fatal1ty from the Aorus competitor.

In short, the results today are very close across all test platforms. Assigning percentages to help visualize normalized performance is both helpful and misleading at times. The ASRock Fatal1ty X399 Professional Gaming performs well in synthetic benchmarks and excellent in all application metrics, but consistently lags behind by a few frames across all gaming tests which ultimately drive its score down. If we compare against the Taichi’s results as well, we clearly see that this board design is an excellent performer, regardless of what extra features are added or removed from the final product.

MORE: Best Motherboards

MORE: All Motherboard Content

Overclocking – Memory, CPU, and Threadripper 2 Results

Initial Overclocking

Having access to only 4 sticks of DDR4-3200 for our review rig, we are forced to populate only a single stick into each of the board’s channels. We’ve successfully run these sticks at DDR4-3333 on other X399 samples, so it is disappointing that this sample is only capable of DDR4-3133. The system will boot with 3200MT/s, but AIDA64 brings the system to a staggering halt. With access to SOC loadline calibration and finer levels of DRAM tuning, this board likely can reach DDR4-3333 or potentially higher.

Our 1950X Threadripper has proven to be fairly straightforward in the overclocking front, but we wanted to see how far we could push it this time around. Unlike our other reviews, at the 3900MHz setting, we did have to bump up our Vcore to a plus 75mV offset for stable Prime95 runs. That 75mV also buys us 50MHz before additional voltage is required. As we increase the voltages, we are able to hit higher clock rates at the cost of thermal throttling. Playing the hot-potato voltage-frequency game, we can reach 4GHz. But throttling continues to mess with our mojo, so we decided to take off our gloves and dig in deeper.

This board has a switch for disabling PROCHOT, which is useful for bypassing throttling governed by the processor. So let’s turn that off. Dr. MOS also has features for disabling over-current and over-voltage protection mechanisms, so we switched those off too. We continued to see throttling events occurring, so we hopped into Ryzen Master and disabled PROCHOT there too (which is NOT consistent on multiple reboots). We also tinkered with the cTDP and other NBIO options in the UEFI and were are able to reduce throttling a bit, but we think we have reached the thermal capabilities of our cooler, running roughly 86C Tdie or 113C Tctl. Speaking of thermal sensors, be cautious trusting the HWiNFO 64’s display on the VRM temperatures. We observed a behavior where once the VRM temperature hits 102C the sensor bugs out and either reports 0C or some negative temperature. If using this board for extreme overclocking, be sure to implement thermocouples or other measurement reporting tools.

32 Core Testing: Meet the Professional Gamer

While wrapping up this review, we got hour hands on a new Threadripper 2990WX processor sample, so we decided to toss it into the Professional Gamer and see how well it does with our full test suite. For our purposes here today, we're only concerned with a high level "what does a 32-core processor look like in our test suite, " and seeing how well it performs from a thermal/overclocking perspective. For a more thorough analysis of the 2990WX, head over to our full review of the Threadripper 2990WX.

Performance wise, this 32-core beast's strengths are synthetic workloads and multi-threaded applications, such as our Sandra suite and Blender tests. There we see up to an 86 percent performance increases compared to our 1950X processor. Unfortunately, even the XFR boost frequencies can't keep up with other synthetics that rely on core speed rather than core counts, such as Adobe and Handbrake. We also observe that the increased die counts impact memory bandwidth and various other architectural-based operations such as encryption.

Gaming is clearly not the target market for this processor, but the ASRock X399 Professional Gaming is able to utilize the 2000 series' features to help mitigate some of the frequency issues for these processors. Across the board, the 2990WX hits performance gaps of upwards of 30 percent from the average, though increasing the resolution and detail settings help shift the burden to the GPU. We also start to see CPU bottlenecking our Talos benchmark, where regardless of detail settings we see consistent frame rates from the 2990WX.

Our power data comes as no surprise, as this 2990W processor clearly draws more power than our GPU, with Small FFT Prime95 running, thus making our efficiency charts laughable. At the end of the day, the overall performance of this board and processor combination should not surprise anyone, and the combined performance is within 6 percentage points of the 1950X setup.

Now, back to this page's topic: overclocking. Pulling 250W by definition already limits processors and motherboards on their ability to overclock, and we instantly hit these constraints when starting our attack run. XFR and the ASRock design do a great job at boosting the processor up to the thermal limit of 67C Tdie, and manually increasing the frequency up to 3275MHz hits our predefined thermal and voltage limits for Prime95 runs. We have seen people claim to overclocked upwards of 4GHz with this processor. At first, we were stumped, but those claims included other cooling solutions often implement much lighter workloads.

We repurpose our Cinebench R15 run to loop 10 instances of the multi-CPU benchmark. And sure enough, we hit an overclock of 3925MHz while hitting VReg thermals in the 115C range and Tdies in the 86C range. \Clearly, this board's Vreg can't cool the 850W of system draw with this test. But for sporadic full-load scenarios, impressive clocks can be achieved.

For users who want a more turn-key approach to overclocking, PBO is a decent alternative, but that comes at a cost. The motherboard and firmware have a delicate balance to achieve. And to do that, larger voltages are set that this builder is not comfortable with for 24/7 use cases. However, these settings do manage to achieve modest burst-load scenarios such as our single run Cinebench tests, but quickly throttle clock rates back to maintain the thermal and power limits deemed safe by the UEFI.

Our Cinebench chart shows that a traditional overclock provides safer voltage control as well as higher maximum and average Cinebench runs compared against the one-button-click solution. Again, better cooling solutions are required to adequately unleash the 32-core beast, and we hope to address that in future content coming your way.

So, what does this all mean to a professional gamer? A decent AIO will grant you modest overclocks compared to other boards, and the regulators are capable of high levels of prolonged stress cases on 190W processors. As for the 2000 series Threadrippers, we hit results similar to ASRock's experiences, and they were able to achieve very good Cinebench scores while using their ASRock X399M Taichi platform. Needless to say, the VReg on this board is capable of juicing up the 32-core monsters to impressive levels while employing the same VReg cooler and MOSFET drivers as both Taichi variants--as long as the load is within reason.

MORE: Best Motherboards

MORE: All Motherboard Content

Temps, Efficiency, Value & Conclusion

Power, Temperatures And Efficiency

As we mention in our Gigabyte X399 Designaire EX review obtaining a maximum wattage power measurement for a system can be an art. Running idle, CPU- and GPU-specific tests are easy to get consistent results. But capturing instances of 100 percent utilization at 100 percent TDP can be difficult with resource sharing occurring across the system. With that said, the ASRock Fatal1ty X399 Professional Gaming motherboard scores well at idle and GPU load conditions, but does pull an extra seven watts compared to the Gigabyte X399 Aorus Gaming 7. Also, that seven watts turns into 10 watts at full load, while the Taichi and MSI boards boast impressive full-system load numbers. As with the performance section, the Taichi and Professional Gaming sport nearly identical PCBs. So for our purposes, we will only factor in idle, CPU, and GPU load numbers.

Our Fractal Design Celdius S36 cooler has remained firmly in the intake position for our Corsair 570X chassis after our launch article, and it’s clear that this position in this chassis is the preferred deployment. Coming in at 36.5C above ambient at full CPU load, the Fatal1ty reports the best CPU temperatures we’ve recorded in this section of the test. Also of note is that the VReg temperatures are noticeably warmer than the Taichi we tested in 2017, despite the same regulator implementation.

It should come as no surprise to anyone right now that such tight average performance metrics will lead to skewed efficiency charts due to overall power. Given the disparity in the total system stress values, we will be focusing on idle, CPU, and GPU metrics alone for efficiency purposes today.

Less than 0.8 percent separates all of the boards’ performance figures, so both ASRock boards are representing positive efficiency results. The Gigabyte and MSI boards both suffer from roughly 2 percent extra power draw, forcing their efficiency figures to be negative. It is good to see that ASRock continues to design and produce a product that is efficient, given the high horsepower available for the system.

Value & Conclusion

Pricing is a critical balance for any vendor or retailer who wants to get stock off shelves and keep employees busy making new products. With the high cost of entry into the HEDT space, expect to pay the cover charge. Clearly, the ASRock X399 Fatal1ty Professional Gaming tops the MSRP charts around $450 at the time of writing, which is nearly $60 more expensive than competing boards. Even against its nearly identical sibling, the Taichi is about 25 percent less expensive.

But for high-end workstations, 10Gb is becoming a necessity. Having access to high-bandwidth network solutions without sacrificing PCIe slots or additional drives in a system make the price premium an excellent value proposition for any business needing high-bandwidth network connectivity. Though single-port 10Gb NICs are dropping in price (at least on eBay), not having to deal with add-on cards helps decrease bill-of-material needs for any larger scale network.

Luckily, we live in a world of sales. Despite the high MSRP, this board can be found for as low as $390, whereas less-expensive boards still linger in the $320 realm. Maybe the introduction of newer X399 boards will help drive prices down for builders.

As hyped as we are about this board, we feel the Fatal1ty and Taichi products should have swapped designs. Gamers don’t need 10Gb controllers, especially when they have to spend the extra money for the luxury. The Taichi board feels to more “professional” aesthetically, and likely could have grabbed a few willing wallets to pay the extra fee for the high-speed Ethernet. Both boards perform well and can overclock well in the right hands, but the 10Gb inclusion just isn’t enough to win over gamers--or to win any awards this time around.

MORE: Best Motherboards

MORE: All Motherboard Content

AMD Ryzen Threadripper 2950X Review: Striking The Balance

palcorn@outlook.com (Paul Alcorn) — Fri, 28 Sep 2018 13:00:00 +0000

The Refreshing Refresh

AMD's 32-core, 64-thread Ryzen Threadripper 2990WX gives high-end desktop users access to the most compute horsepower available from a single CPU socket. But while it may be on many enthusiasts' wish lists, the processor's unique architecture causes poor performance in many common desktop applications. Moreover, an $1800 price tag makes the flagship Threadripper a niche product, even among professionals accustomed to paying a premium for workstation hardware.

One of the best CPUS for desktop applications, the Ryzen Threadripper 2950X is unquestionably a better value proposition for the masses, offering 16 cores and 32 threads at a $900 price point. AMD's only real problem is that its own previous-gen Threadripper chips sell for less: the 16C/32T Threadripper 1950X can be found for $700, while the 12C/24T is available for under $500.

So, should you spring for the 2950X and its 12nm transistors, lower memory/cache latency, higher clock rates, and enhanced multi-core Precision Boost, or compromise a bit by buying an older Threadripper chip before they disappear for good? The 2950X's features do deliver tangible performance improvements over previous-gen Threadripper models, meaning you do get a lot of bang for your buck.

Ryzen Threadripper X-Series

AMD split its Threadripper family up into the WX and X series. The former mows through intense multitasking, software development, video/audio production, and content creation. The latter is aimed at gamers and prosumers.

	Ryzen Threadripper 2990WX	Ryzen Threadripper 2950X
Socket	TR4	TR4
Cores / Threads	32 / 64	16 / 32
Base Frequency	3.0 GHz	3.5 GHz
Boost Frequency	4.2 GHz	4.4 GHz
Memory Speed	DDR4-2933 (Varies)	DDR4-2933 (Varies)
Memory Controller	Quad-Channel	Quad-Channel
Unlocked Multiplier	Yes	Yes
PCIe Lanes	64 (Four to the chipset)	64 (Four to the chipset)
Integrated Graphics	No	No
Cache (L2 / L3)	80MB	40MB
Architecture	Zen+	Zen+
Process	12nm LP GloFo	12nm LP GloFo
TDP	250W	180W

Although the $900 Threadripper 2950X "only" offers 16 cores and 32 threads, it serves up much higher clock rates than the 64-thread 2990WX. The 2950X starts with a base frequency of 3.5 GHz and boosts up to 4.4 GHz (a slight step up from the previous-gen Threadripper 1950X's 3.4/4.2 GHz). Moreover, the Zen+ architectural enhancements serve up much better benchmark results across a range of workloads compared to AMD's earliest Threadripper models.

All of the 2000-series Threadripper processors are backward-compatible with existing X399 motherboards. That's good news given the high prices on those platforms. While older Socket TR4-equipped boards may struggle under the power requirements of AMD's 250W Threadripper 2990WX and 2970WX, particularly if you try to overclock, those same motherboards were designed to accommodate (and overclock) the older Threadripper 1950X flagship. As a result, existing X399 platforms should have enough headroom to enable most of the 2950X's Precision Boost Overdrive capabilities for higher frequencies when they're needed.

	Cores /Threads	Base / Boost (GHz)	L3 Cache (MB)	PCIe 3.0	DRAM	TDP	MSRP	Price Per Core
TR 2990WX	32 / 64	3.0 / 4.2	64	64 (4 to PCH)	Quad DDR4-2933	250W	$1799	$56
TR 2970WX	24 / 48	3.0 / 3.2	64	64 (4 to PCH)	Quad DDR4-2933	250W	$1299	$54
Core i9-7980XE	18 / 36	2.6 / 4.4	24.75	44	Quad DDR4-2666	140W	$1999	$111
TR 2950X	16 / 32	3.5 / 4.4	32	64 (4 to PCH)	Quad DDR4-2933	180W	$899	$56
TR 1950X	16 / 32	3.4 / 4.4	64	64 (4 to PCH)	Quad DDR4-2667	180W	$750	$47
Core i9-7960X	16 / 32	2.8 / 4.4	22	44	Quad DDR4-2666	140W	$1699	$106
TR 2920X	12 / 24	3.5 / 4.3	32	64 (4 to PCH)	Quad DDR4-2933	180W	$649	$54
TR 1920X	12 / 24	3.5 / 4.2	64	64 (4 to PCH)	Quad DDR4-2667	180W	$399	$33
Core i9-7920X	12 /24	2.9 / 4.4	16.50	44	Quad DDR4-2666	140W	$1199	$100
Core i9-7900X	10 / 20	3.3 / 4.3	13.75	44	Quad DDR4-2666	140W	$999	$99
Core i7-8700K	6 / 12	3.7 / 4.7	12	16	Dual DDR4-2666	95W	$359	$60
Ryzen 7 2700X	8 / 16	3.7 / 4.3	16	16	Dual DDR4-2933	105W	$329	$41

We detailed the second-gen Threadripper architecture in our review of the 2990WX. In short, though, Ryzen Threadripper 2950X mirrors the layout of AMD's first-gen Threadripper chips: two Zeppelin dies are connected via another layer of the Infinity Fabric. AMD flanks them with a pair of dummy dies that serve as non-functional fillers, ensuring the heat spreader's structural integrity and consistent mating with the socket's pins. This configuration demonstrates the same eccentricities as AMD's previous models, which are largely borne of the multi-chip design. Fortunately, the company's architectural improvements do soften the impact in workloads that were more severely affected last generation.

AMD ships all Threadripper CPUs with an Asetek bracket that provides partial coverage of the massive heat spreader using certain closed-loop liquid coolers. According to AMD, this partial coverage is fine for stock operation. But we found that full-coverage coolers work better. AMD also collaborated with Cooler Master to develop the Wraith Ripper heat sink/fan combo for its Socket TR4 interface. It's sold separately, though.

As per usual, AMD uses Indium solder between its dies and heat spreader to improve thermal transfer. In contrast, Intel employs thermal grease on its highest-end processors. Intel also recommends liquid cooling for its Skylake-X processors. AMD says that's not necessary for Threadripper.

DIMM Config	Memory Ranks	Official Supported Transfer Rate (MT/s)
4 of 4	Single	DDR4-2933
4 of 8		DDR4-2667
8 of 8		DDR4-2133
4 of 4	Dual	DDR4-2933
4 of 8		DDR4-2667
8 of 8		DDR4-1866

All of the new Threadripper chips come equipped with the hallmarks of AMD's Ryzen value proposition, such as unlocked ratio multipliers for overclocking and 60 lanes of third-gen PCI Express (plus four lanes attached to the supporting chipset). Copious connectivity could come in handy for multiple add-in graphics cards, but it's also useful for high-performance storage and networking.

Threadripper CPUs feature independent dual-channel memory controllers located on two dies, which combine to provide quad-channel support with varying data transfer rates based upon your configuration. With the second-gen Threadripper processors, AMD bumps its maximum specification to DDR4-2933 (up from DDR4-2666).

The platform supports ECC memory and up to 256GB of capacity, but it can accommodate up to 2TB as density increases. We've already seen new, denser DRAM coming from the likes of Samsung, making support for more capacious memory configurations a future-looking feature.

MORE: Best CPUs

MORE: All CPUs Content

Overclocking & Test Setup

Overclocking AMD's Ryzen processors is a fairly straightforward process. Most first-gen Threadripper 1950X CPUs tap out between 3.9 and 4.0 GHz. However, our 2950X maintained 4.1 GHz with ease. We didn't run into serious thermal constraints during our testing, but we do recommend water cooling (preferably with a full-contact block) for overclocking. The included Asetek mounting bracket is sufficient, but this partial-coverage solution isn't as robust as our full-coverage Enermax Liqtech 240 TR4 II cooler.

We dialed in a 4.1 GHz overclock by tuning the Vcore to 1.375V and bumping VDD_SoC up to 1.2V. We also increased the memory to DDR4-3466 with 14-14-14-34 timings and assigned the default LLC setting.

We recorded 246W of power consumption during our torture test with Prime95 and the processor at 4.1 GHz in Creator Mode. AMD's Precision Boost Overdrive serves much the same purpose as our manual overclock, but with clock rates that adjust dynamically to meet demand. Our Threadripper 2950X sample sustained an all-core 4.1 GHz with PBO active, drawing a steady 250W under load.

With PBO active, the CPU drops to lower clock rates during idle periods. Our Threadripper 2950X fell to 2.7 GHz, resulting in a ~31W power measurement. In contrast, our all-core overclock drew 36W at idle, illustrating the efficiency benefit of using AMD's automatic overclock feature rather than dialing in a manual adjustment.

Comparison Products

Intel Core i9-7980XE

Intel Core i9-7960X

Intel Core i9-7900X

Test Setup

We tested the Threadripper 2 models with MSI's MEG X399 Creation motherboard. Due to cooling and power delivery constraints, we ran through our full test suite at stock settings, with PBO activated, and with an all-core 4.1 GHz overclock. Our PBO-enabled configurations did benefit from higher memory transfer rates, as detailed in the table below.

Test System & Configuration
Hardware	Germany AMD Socket AM4 (400-Series)AMD Ryzen 7 and Ryzen 5 MSI X470 Gaming M7 AC 2x 8GB G.Skill FlareX DDR4-3200 @ DDR4-2667, DDR4-3466AMD Socket SP3 (TR4)Threadripper 2MSI MEG X399 Creation 4x 8GB G.Skill FlareX DDR4-3200 RGBIntel LGA 1151 (Z370): Intel Core i7-8700K MSI Z370 Gaming Pro Carbon AC 2x 8GB G.Skill FlareX DDR4-3200 @ DDR4-2667, DDR4-3466Intel LGA 2066 Intel Core i7, Core i9 MSI X299 Gaming Pro Carbon AC 4x 8GB G.Skill FlareX DDR4-3200 @ DDR4-2666All SystemsGeForce GTX 1080 Founders Edition (Gaming) Nvidia Quadro P6000 (Workstation)1x 1TB Toshiba OCZ RD400 (M.2, System SSD) 4x 1TB Crucial MX300 (Storage, Images)be quiet! Dark Power Pro 11, 850W Windows 10 Pro (All Updates)U.S. AMD Socket SP3 (TR4)Threadripper Gen 1 & 2MSI MEG X399 Creation 4x 8GB G.Skill FlareX DDR4-3200 @ DDR4-2933, DDR4-3200, DDR4-3466Intel LGA 2066Intel Core i9-7960X, -7980XE, -7900XMSI X299 Gaming Pro Carbon AC4x 8GB G.Skill FlareX DDR4-3200 @ DDR4-2666, DDR4-3200AMD Socket AM4 (400-Series)AMD Ryzen 7 2700X MSI X470 Gaming M7 AC2x 8GB G.Skill FlareX DDR4-3200 @ DDR4-2933Intel LGA 1151 (Z370)Intel Core i7-8086K, Core i7-8700K, Core i5-8600K, Core i5-8400, Core i7-8700MSI Z370 Gaming Pro Carbon AC2x 8GB G.Skill FlareX DDR4-3200 @ DDR4-2667All EVGA GeForce GTX 1080 FE 1TB Samsung PM863SilverStone ST1500-TI, 1500WWindows 10 Pro (All Updates)
Cooling	GermanyAMD Wraith RipperAlphacool Ice Block XPXEnermax LiqTech 240 TR4Thermal Grizzly KryonautU.S.Wraith RipperCorsair H115iEnermax Liqtech 240 TR4 II
Power Consumption Measurement	Contact-free DC Measurement at PCIe Slot (Using a Riser Card) Contact-free DC Measurement at External Auxiliary Power Supply Cable Direct Voltage Measurement at Power Supply 2x Rohde & Schwarz HMO 3054, 500 MHz Digital Multi-Channel Oscilloscope with Storage Function4x Rohde & Schwarz HZO50 Current Probe (1mA - 30A, 100 kHz, DC) 4x Rohde & Schwarz HZ355 (10:1 Probes, 500 MHz) 1x Rohde & Schwarz HMC 8012 Digital Multimeter with Storage Function
Thermal Measurement	1x Optris PI640 80 Hz Infrared Camera + PI Connect Real-Time Infrared Monitoring and Recording
Acoustic Measurement	NTI Audio M2211 (with Calibration File, Low Cut at 50Hz) Steinberg UR12 (with Phantom Power for Microphones)Creative X7, Smaart v.7 Custom-Made Proprietary Measurement Chamber, 3.5 x 1.8 x 2.2m (L x D x H) Perpendicular to Center of Noise Source(s), Measurement Distance of 50cm Noise Level in dB(A) (Slow), Real-time Frequency Analyzer (RTA) Graphical Frequency Spectrum of Noise

MORE: Best CPUs

MORE: All CPUs Content

Test Notes

Unlike the first-gen Ryzen Threadripper 1950X, AMD aims its 2950X at enthusiasts and gamers. So, for this review, we tested Threadripper 2950X using AMD's Game Mode setting. Just bear in mind that the company also provides toggles that allow you to customize settings for individual applications.

We tested across our gaming suite using a 1920x1080 resolution, minimizing graphics bottlenecks. Of course, as you step up to 2560x1440 or 3840x2160, the difference between processors shrinks.

VRMark, 3DMark

We aren't big fans of using synthetic benchmarks to measure performance, but 3DMark's DX11 and DX12 CPU tests provide useful insight into the amount of horsepower available to game engines.

During the DX11 and DX12 tests, automated overclocking through Precision Boost Overdrive (PBO) yields very similar performance to our all-core 4.1 GHz adjustment. Both settings give us a nice boost over stock performance, and there is very little differentiation between the two tuning options. Interestingly, using PBO on the MSI X399 Creation often results in an all-core 4.1 GHz clock rate, albeit with sporadic jumps to 4.2 GHz on a few cores.

UL's VRMark test lets you gauge your system's suitability for use with the HTC Vive or Oculus Rift, even if you don't currently own an HMD. UL defines a passing score as anything above 109 FPS.

This lightly threaded workload allows Threadripper 2950X's PBO feature to trigger a 4.4 GHz boost frequency, which we'd expect to beat our static 4.1 GHz all-core overclock.

Ashes of the Singularity: Escalation

Ashes of the Singularity: Escalation is a computationally intense title that scales well with thread count. PBO allows Threadripper 2950X to stay in its boost states for longer periods of time, though the CPU still down-clocks under some conditions.

MORE: Best CPUs

MORE: All CPUs Content

Civilization VI AI Test

Civilization VI's AI test measures CPU performance in a turn-based strategy game and tends to favor per-core performance.

Our 4.1 GHz manually overclocked 2950X beat the PBO-enabled configuration by a slim margin, though even a stock 2950X outperformed AMD's tuned Threadripper 1950X.

Civilization VI Graphics Test

A nice mixture of high clock rates and parallelism helped AMD in the Civilization VI graphics test. Threadripper 2950X profited handsomely from its PBO feature, while the 4.1 GHz overclock trailed AMD's stock configuration. That shouldn't be surprising, though: the 2950X has a 4.4 GHz boost frequency right out of the box.

The CPU benefited from its lofty ceiling during bursty activity, and this benchmark's 25-second run time allowed the 2950X to stay near its sweet spot.

Warhammer 40,000: Dawn of War III

The 4.1 GHz all-core overclock effectively tied our configuration with PBO active. But the real story was Threadripper 2950X's nice speed-up compared to AMD's previous-gen Threadripper 1950X. The stock 2950X beat its predecessor by an average of 9 FPS, which is no small feat in this title. Intel's Core i7-8700K and tuned Core i9-7960X still dominated up top, though.

MORE: Best CPUs

MORE: All CPUs Content

Far Cry 5

All three Core i9 processors encountered a sharp drop-off in our 99.9th percentile metrics, emphasizing a handful of hiccups during the benchmark. Overclocking helped smooth those out.

Meanwhile, Threadripper 2950X served up a smoother frame rate profile.

Grand Theft Auto V

Grand Theft Auto V favors Intel architectures and, more generally, multi-core designs with high clock rates.

This title prefers the higher boost clock rates available through PBO. Though Intel led through our benchmark, it is clear that AMD's targeted improvements made Threadripper 2950X a much more agile competitor than the previous-gen model.

Hitman

Our Hitman benchmark was rendered almost useless by a patch that imposed a 90 FPS performance cap. A subsequent update restored our test to its prior glory.

The 0.3 FPS difference between our all-core overclock and PBO turned on fell within a margin of error.

MORE: Best CPUs

MORE: All CPUs Content

Middle-earth: Shadow Of War

Shadow of War leans heavier on graphics resources than pure host processing, so we didn't see large deltas between the fastest and slowest CPUs. However, we also spotted a repeatable period of inconsistent performance from the Core i9-7980XE that reminded us average frame rates rarely tell the entire story.

Project CARS 2

Although Project CARS 2 is purportedly optimized for threading, clock rates obviously affect frame rates. The Core i9 and i7 processors led through our benchmark, though AMD's Ryzen Threadripper 2950X was surprisingly competitive. It even outperformed the mainstream Ryzen 7 2700X.

MORE: Best CPUs

MORE: All CPUs Content

Adobe Creative Cloud

Even though this suite has a few parallelized workloads, the final score is heavily influenced by the lightly-threaded tasks common in most desktop applications.

AMD's Threadripper 2950X offers a well-balanced feature set. But the 2990WX’s strength in the heavily threaded InDesign and Photoshop Heavy tests allowed it to capture an overall lead. The tuned 1920X regularly popped up ahead of the 2950X, and we verified the outcome through extensive retesting.

Web Browser

The Krakken suite evaluates JavaScript performance using several workloads, including audio, imaging, and cryptography. Like most web browser workloads, single-threaded performance reigns supreme. These tests expose the trade-offs you make for an all-core overclock.

AMD's second-gen Threadripper line-up goes a long way to improve the performance of lightly threaded workloads. But the value of its dynamic overclocking feature can't be understated. As we saw in the MotionMark benchmarks, which emphasize graphics (rather than JavaScript) and are also sensitive to CPU clock rates, PBO offered a big performance boost with minimal effort on our part (aside from the investment in a capable cooler).

Productivity

AMD's Ryzen Threadripper 2990WX suffered due to its low clock rates, and PBO did little to help. Unfortunately, manual tuning of the 2990WX really isn't viable given its hefty core count and resulting power consumption. Meanwhile, the Threadripper 2950X chipped away at Intel's application loading prowess, though we don't expect high-end desktop CPUs to usurp mainstream alternatives like the Ryzen 7 2700X and Core i7-8700K in this test.

Our video conferencing suite measures performance in single- and multi-user applications that utilize the Windows Media Foundation for playback and encoding. It also performs facial detection to model real-world usage. Threadripper 2950X's processing time showed little benefit from one setup to the next, though.

The photo editing benchmark measures performance with Futuremark's binaries using the ImageMagick library. Common photo processing workloads also tend to be parallelized. Nevertheless, Threadripper 2950X suffered tremendously from our 4.1 GHz all-core overclock, demonstrating sensitivity to frequency as well. At least AMD's dynamic overclocking feature boosted performance into the same league as the brawnier 32C/64T 2990WX.

MORE: Best CPUs

MORE: All CPUs Content

Rendering

Many of these workloads stress the memory subsystem, diminishing Ryzen Threadripper 2990WX's big core count advantage due to accesses from the remote memory controllers.

Threadripper 2950X doesn't suffer the same fate. Rather, it chews through some of the challenging tasks that stymied AMD's first-gen 1950X. In fact, the 2950X in stock form often beats or lands close to the overclocked 1950X.

Intel’s processors maintain their lead in the single-threaded POV-Ray and Cinebench tests, but it's easy to see that AMD’s extra cores help offset their lower IPC in threaded benchmarks.

Encoding & Compression

Our threaded compression and decompression metrics work directly from system memory, removing storage throughput from the equation. This workload benefits heavily from threading, but either memory throughput or poor software scaling holds the 29990WX back from realizing
its potential in the compression test. Threadripper 2950X, which utilizes two dies with directly-attached memory controllers, offers a nice boost compared to the 1950X at stock settings. Surprisingly, the 2950X's 4.1 GHz all-core overclock is faster than the same chip with PBO enabled, even though PBO offers the benefit of higher boost frequencies. Closer examination of our benchmark reveals that, even though the workload stresses all cores heavily, it is sporadic in nature. That forces the 2950X to frequently adjust its clock rate, and like all adaptive algorithms, slow response times can penalize performance.

y-cruncher, a single- and multi-threaded program that computes pi using AVX instructions, is a great test to measure Threadripper’s AVX performance. Intel’s Core i9 employs two 256-bit AVX FMA units per core that operate in parallel, whereas Ryzen's Zen architecture divides 256-bit AVX operations across two FMA units per core. Intel's AVX instruction support shines during the single-threaded benchmark. However, spreading the workload across the 2950X's 16 cores and 32 threads puts it on competitive footing.

Threadripper 2950X outstrips the much more expensive 2990WX in the HandBrake x265 test, which relies heavily on AVX instructions, and the H.264 test.

MORE: Best CPUs

MORE: All CPUs Content

Final Analysis

There are plenty of reasons not to buy AMD's Ryzen Threadripper 2990WX. Sure, 32 cores and 64 threads sound like a dream come true for prosumers and professionals. But the CPU comes with a list of caveats. You really need a high-end motherboard, a high-end power supply, and high-end cooling to extract maximum performance from your $2000 investment. Some threaded applications simply don't scale well across that many cores. When you factor in some of the odd performance results we've seen from other applications, it becomes necessary to exercise caution before taking the leap. Make sure your applications can properly utilize the 2990WX's available resources, and that you're willing to make some compromises elsewhere.

In contrast, Threadripper 2950X doesn't impose such stringent motherboard, power supply, and cooling requirements. It consumes a little more power than its predecessor at stock and overclocked settings, but the difference isn't large enough to affect previous-generation motherboards designed for overclocking. You certainly don't have to worry about pulling up to 500W through the socket under load like we saw from Threadripper 2990WX.

Considering the raw horsepower we're talking about, AMD's pricing is extremely competitive. Ryzen Threadripper 2950X kept pace with the $1700 Core i9-7960X in many of our tests, but sells for almost half of its price. Unfortunately, the X399 platform remains prohibitively expensive. And populating all four memory channels also gets expensive in light of today's egregious DRAM pricing.

AMD's Precision Boost Overdrive is the unsung hero of the second-gen Threadripper launch. Any enthusiast can attest to the benefits of overclocking, but tuning isn't for everyone. AMD's technology brings thought-free overclocking to the masses, and unlike some of the dynamic implementations we've seen on Intel motherboards, it doesn't pump in a ridiculous amount of voltage to the processor. The adaptive algorithms do a great job of responding to the capabilities of your specific system, meaning you can extract more value from beefier components.

While we still recommend the mainstream Ryzen 7 2700X or Core i7-8700K for gaming, they clearly can't keep pace with Threadripper in productivity-oriented applications. Intel's Skylake-X processors are still brutally fast, but you'll also pay a premium for the privilege of owning one. The Threadripper 2950X offers a lot more performance at a lower price than the first-gen Threadripper did at launch, but we wouldn't recommend a direct upgrade from the 1950X. If you're looking to upgrade from an older CPU to this well-balanced processor, AMD's Ryzen Threadripper 2950X does not disappoint.

MORE: Best CPUs

MORE: All CPUs Content

First Generation Ryzen Threadripper Prices Plummet

Zhiye Liu — Mon, 27 Aug 2018 18:02:02 +0000

Now that the Ryzen Threadripper 2000-series processors have officially launched, AMD is issuing significant price cuts on its three first-generation Threadripper processors.

Despite being a year old, AMD's Ryzen Threadripper 1000-series processors are formidable multi-core beasts for consumers that constantly deal with heavy-threaded workloads. The Threadripper 1950X, which was the flagship processor, originally launched at $999. The processor now has a suggested retail price (SEP) of $799. In fact, at the time of writing the Threadripper 1950X is available on Amazon for $718.90, which is a pretty sweet deal for a processor that comes equipped with 16 cores and 32 threads.

AMD is also lowering the price for the Threadripper 1920X and Threadripper 1900X models as well. The chipmaker has adjusted the suggested retail price for the Threadripper 1920X with 12 cores and 24 threads from $799 to $399 and the Threadripper 1900X with eight cores and 16 threads from $549 to $299. At the time of this article, the Threadripper 1920X's pricing on Amazon doesn't reflect the price change and it's still priced at $463.06; however, the Threadripper 1900X is selling for $307.

Whether you're interested in picking up the monstrous 16-core Threadripper 1950X or the entry-level octa-core Threadripper 1900X, it's important to pair it with a capable X399-based motherboard and CPU cooling solution to exploit the processor's full potential.

Model	Threadripper 1950X	Threadripper 1920X	Threadripper 1900X
Suggested Etail Price	$799	$399	$299
Launch Price	$999	$799	$549
Interface / Chipset	TR4 / X399	TR4 / X399	TR4 / X399
Cores / Threads	16 / 32	12 / 24	8 / 16
TDP	180W	180W	180W
Base Frequency (GHz)	3.4	3.5	3.8
Boost Frequency (GHz)	4.0 (4.2 XFR)	4.0 (4.2 XFR)	4.0 (4.2 XFR)
Cache (L2+L3)	40MB	38MB	20MB
Memory Support	DDR4-2667	DDR4-2667	DDR4-2667
Memory Controller	Quad-Channel	Quad-Channel	Quad-Channel
Unlocked Multiplier	Yes	Yes	Yes
PCIe Lanes	64	64	64

Gigabyte X399 Designare EX Motherboard Review

Jacob Terkelsen — Sun, 19 Aug 2018 13:00:00 +0000

Features & Layout

We'll just come right out and say it: The Gigabyte X399 Designare EX is the most beautiful board we’ve tested from the AMD camp thus far. It's also a top performer in our charts, but comes with a price tag to match--it's currently selling for between $370 and $400. Builders focused on value should obviously look elsewhere (though no Threadripper boards come cheap). But builds that require top-tier performance and stand-out aesthetics will find the Designare appealing.

We really liked the Gigabyte X399 Aorus Gaming 7, so when Gigabyte initially posted beauty shots of the Designare variant on Instagram, our jaws dropped. Instantly, we were envisioning how well it would look in our Corsair Crystal 570X lit up for the world to see. But, we've been burned before when judging a product solely by press shots. So we'll proceed with cautious optimism.

Specifications

Socket	TR4
Chipset	AMD X399
Form Factor	ATX
Voltage Regulator	8+3 Phases
Video Ports	✗
USB Ports	10Gbps:(1) Type A, (1) Type C5Gb/s: (8) Type A
Network Jacks	(2) Gigabit Ethernet
Audio Jacks	(5) Analog, (1) Digital
Legacy Ports/Jacks	(1) PS/2
Other Ports/Jack	(2) MMCX WiFi Antenna
PCIe x16	(4) v3.0 (x16/x8/x16/x8)(1) v2.0 (x4)
PCIe x8	✗
PCIe x4	✗
PCIe x1	✗
CrossFire/SLI	4x / 4x
DIMM slots	(8) DDR4
M.2 slots	(2) PCIe v3 x4 / SATA3 60/80/110mm(1) PCIe v3 x4 / SATA3 42/60/80mm
U.2 Ports	M.2 Adapter
SATA Ports	(8) 6Gb/s
USB Headers	(1) 5 Gbps Type-A(1) 5 Gbps Type-C(2) USB2.0
Fan Headers	(8) 4-Pin [1 is dedicated to Vreg fan]
Legacy Interfaces	✗
Other Interfaces	(3) RGB-LED, (2) Temp Sensor, TPM, Clear CMOS
Diagnostics Panel	Numeric
Internal Button/Switch	CMOS Clear, Power, Reset
SATA Controllers	Integrated (0/1/10)
Ethernet Controllers	(2) Intel® I211AT
Wi-Fi / Bluetooth	Intel Dual Band Wireless-AC 8265 cardBluetooth 4.2 / 3.0
USB Controllers	✗
HD Audio Codec	ALC1220
DDL/DTS Connect	✗ / ✗
Warranty	3 Years

The Gigabyte X399 Designare EX shares the same design thinking as some of its Intel-based counterparts, which is a pleasant departure from the Aorus brand’s emphasis on LEDs and angled, feathery icons. Favoring elegant, sleek lines and a blue-grey color palette help elevate these boards from workhorses to works of art. Comparing against the Intel X299 variant, we see similar IO shielding, M.2 heatsinks, and regulator and south bridge heatsink designs.

But for the AMD variant, the full Threadripper assortment of connectivity is provided, and every inch of planar real estate is utilized for components. Heading to the backpanel, USB headers consume the majority of the integrated backplate. You'll find eight USB 3.0 and two USB3.1 Gen2 connectors (one Type-A and one Type-B). Network connectivity is provided with the Designare through the smaller MMCX antenna connections and dual RJ-45 connectors, enabling 802.11 a/b/g/n/ac and dual Gigabit connections respectively. As for audio, five gold-plated analog ports and an optical S/PDIF connector are available to the Realtek ALC1220 codec. And just in case your trackball hails from the 90’s, a single PS/2 port is available for interrupt-driven IO devices.

Storage and PCIE are strong points for Threadripper, and the Designare provides eight angled SATA3 headers, three PCIe Gen3 x4 NVMe M.2 slots, and five PCIe x16 connectors wired out for x16/x8/Gen2 x4/x16/x8. Fortunately, the silkscreen on the board provides clarification of each of these connectors. Our recommended loadout for these slots would be primary GPU in the first slot, followed by a blank x8 slot, a video capture card in the x4, and a 10Gb network card in the bottom x16 slot.

As expected, the bottom edge of the board contains the bulk of the headers for the system, featuring RGB connections, power/CMOS/reset buttons, debug LED, and various USB headers, with details available in the test configuration section of this review. Ample fan connections are provided on this board, with two located on the bottom edge and five on the top right. The eighth advertised system fan header is wired by default to the backpanel I/O fan, which makes the eight-header claim suspect at best. An additional RGB header is located at the top-right of the board and a USB3.1 Gen2 front-panel header is accessible directly below the 24-pin ATX connector.

We haven’t talked about the memory configurations on any of our previous X399 articles, so now is a good chance to cover that. The Designare EX equips the standard quad-channel DDR4 configuration on the high-end TR4 socket, but blesses us with old-school dual-hinged DIMM connectors for that audible “click-click” confirming secure seating of our expensive DIMM sticks. This board also supports both unbuffered ECC and non-ECC DIMMs, which fit our pro-sumer oriented builds well. We just wish we could find some ECC RAM that wasn’t basic PCB green?

Beyond the 24-pin ATX, an 8-pin and 4-pin EPS-12V connector are located in the top-left area of the board, which is split by the heat pipe connecting both stacks of VReg heat sinks. Luckily, there are no fan headers to deal with here. But it does seem odd to have heatpipes sandwiched between very important power delivery components for the already hungry Threadripper CPU. Speaking of regulators, this board is equipped with an 8-phase VCore design, featuring PowIRstage Integrated MOSFET drivers and a PWM driver running in 8+0 mode. Remember that fan on the back IO? The heatpipes used here help transfer the heat to the larger SOC heat sink and pull air across the stack out the back of the backplate. With Threadripper 2, the additional cooling might come in handy when powering the larger 250W SKUs like the 2990WX.

Finally, let’s talk aesthetics. This Gigabyte board only boasts onboard LEDs illuminating the south bridge chip highlighting the Tron-like gem logo. For some, that might not be enough, but for a toned-down workstation build, it's just the right amount. As I’m in the process of building a Threadripper build in a Phanteks P400 with a similar board, it is a very pleasant combination.Adding in an external LED strip, the shadows generated by the southbridge and the M.2 covers are great.

So what's not to like about this board? To be frank, not much. We feel that the contents of the box are a little bit misplaced, favoring form over function. Also, the top left quadrant of the board is incredibly tight, and we would have preferred higher-quality power and reset buttons on the bottom edge of the board. The inclusion of the near full-coverage backplate seems wasted with most cases. So a little bit of creativity is needed for picking the right chassis for a build with this board.

MORE: Best Motherboards

MORE: All Motherboard Content

UEFI & Test Configuration

Not much has changed for this fancy board on the UEFI front as compared to theGigabyte X399 Aorus Gaming 7. Given that other competing board vendors customize their UEFI’s for several products, this Designare should have seen a similar treatment to match the aesthetic differences from its Aorus counterparts.

Looking beyond the surface, both UEFIs perform and operate on par with each other. The default panel presented is the M.I.T. menu, which opens up options for CPU and memory frequencies, voltage and fan settings, and miscellaneous options for the processor and system. Continuing to drive across the high-level tabs, System shows general time and UEFI version dates, the Peripherals tab opens options for the network interface controller and grants access to the original AMD X399 options, and Chipset primarily offers up PCIe and storage options.

Back to the M.I.T. menu: Smart Fan 5 graces us with its customizable fan curve interface and the various sensors that the motherboard firmware monitors. Fortunately, it appears the CPU reading is tracking the proper sensor and not Tctl, and the VRM MOS reading appears to track well with the increased current load of the system. We still recommend running custom fan curves with the 100% fan setting moved to higher temperatures, since fans can get a little rowdy even in the Silent profile.

We'll talk more about the voltage settings specifically for this board in the overclocking section. But for an overview, the Advanced Voltage menu has access to VCore, VSoc, 1.8V, and custom DRAM termination voltage settings. Also, each of the CMOS drivers has integrated protection circuits that can be adjusted from within the UEFI to either protect or unlock your motherboard. Loadline calibration is also available for fine-tuning the regulators' response to load conditions, so that proper voltage and current is supplied to the processor.

Memory tuning is highly visible with the modern UEFI designs and the Gigabyte X399 Designare EX is no exception. Switching from Auto Memory Timing Mode allows users to tune the standard timing control, and advanced timing is available as well if desired. Miscellaneous settings present users with the ability to turn off Cool & Quiet, SVM Mode, Global C-States, and other processor controls that might be helpful for various use cases.

On to the Peripherals menu. Since we tested this board before the Threadripper 2 release, many options in this tab might have changed or others added and removed as applicable, so consult the UEFI documentation for recent updates (the manual is pretty good as well). For our purposes though, many of these options enable lower level function of the Threadripper processors, particularly in the AMD CBS menu's NBIO and Zen submenu. Zen Common Options has built-in overclocking modes that we do not recommend using. Custom Pstates have shown to be effective for some of the XFR2 and Precision Boost Overdrive overclocking methodologies with Ryzen 2000 series processors, so they likely can be exploited for some Threadripper success. We tinkered with the NBIO options quite a bit during overclocking, and they appear to grant access to various power-related limits in place for the processor. More on that to come in the overclocking section.

Test Hardware

The Corsair Crystal Series 570X still houses our Threadripper test bed, and it pairs well with the Gigabyte X399 Designare EX, as far as aesthetics are concerned. It's a looker, just like the board we're pairing it with.

Antec’s 1200W High Current power supply provides the juice to our Gigabyte GTX 1080 G1 Gaming GPU and Ryzen Threadripper 1950X CPU. G.Skill provides the 32GB of DDR4-3200MHz RAM, and Toshiba's 256GB RD400 M.2 NVMe drive serves as our boot drive. Fractal Design’s Celsius S36 360mm water cooler is placed on the intake side of our chassis to provide cooling for our 190W processor.

MORE: Best Motherboards

MORE: All Motherboard Content

Synthetics & Gaming Benchmarks

Comparison Products

ASRock X399 Taichi

Gigabyte X399 Aorus Gaming 7

MSI X399 Gaming Pro Carbon AC

Synthetics & Applications

Having a full set of Creative Mode data now makes it fairly straightforward to identify any basic trends in our hardware setup, so we will spend a little more time with the synthetics today than usual. Our sixteen core 1950X loves Sandra’s benchmarks, with each of the motherboards showing consistent results across the Arithmetic, Multimedia, Cryptography, and Memory Bandwidth. Having said that, we do observe a few intricacies with the data. In the Multimedia workload, the Integer and floating point AVX results seem to round to the nearest 0.0010 GPix/s as we break the 1GPix barrier, which skews the data a bit. Also, in the Cryptography and memory workloads, we see deltas of up to 2 GB/s. That's hardly noticeable, but it does impact our overall performance results.

PCMark08 decides to take Sandra’s consistency and throw it out the window. Though we do our best to keep chipset, GPU, and operating system conditions similar between reviews, we still can’t rule out that some environmental variable is impacting our results in a measurable manner. The Home, Work, Creative, and Office workloads clearly favor the Designare’s environment, with upwards of 11 percentage points separating the next contender. Oddly, Adobe decides to rain on the Designare’s parade and places that board in last place. Nevertheless, Storage is our only reliable workload showing less than 0.11 percent separating all the samples.

Speaking of storage, our IOmeter workload is now a standard sight in our synthetics segment, and each motherboard sample reviewed wins at least one of the individual metrics today. With performance deltas of between and 10.5 percentage points, the Gigabyte X399 Gaming Pro Carbon AC’s squeaks ahead overall with disk performance.

Cinebench R15 still produces consistent results when comparing single and multi threaded scores. With a 3 point win on single-thread and a 14 point lead on the ASRock Taichi, the Gigabyte X399 Aorus Gaming 7 wins this round. Compubench results show the Designare performing well in three of the four graphed workloads, but this board in fact wins in the majority of this more graphics-intensive suite.

As we shift over to the 3DMark workloads, we start to see each of these boards starting to take wins for various metrics, which is a good thing. For Skydiver, the Designare configuration is completely overkill at 720p and rakes in healthy wins for everything but the Physics tests. We still can’t explain the Taichi’s gruesome defeat here, but it does come back with Firestrike and wins all the categories. As for the Designare, a second-place finish with Firestrike continues to slip as its Physics score pulls it to third place in Firestrike Extreme.

Game Benchmarks & Overall Performance

Gaming is not particularly Threadripper's ideal use case, but with the right components it can be a powerhouse of frames, effects, and butter-smooth gameplay. As we’ve experienced with prior reviews, F1 2015 continues struggle with the high core counts of the 1950X, and appears to bind up once we enable more than 24 cores. After tweaking our automation XMLs, the Designare still lags behind by up to 11 frames per second at 1080p and Ultra High detail settings, which translates to nearly a five-percentage-point gap between it and the Aorus Gaming 7. At 4k, the MSI Carbon Pro Gaming AC takes the lead, but the variance between the test samples is hardly noticeable while playing the game.

The Talos Principle shows some interesting trends. At 1080p and High details, the Designare lags by 12 percentage points compared to the MSI, and at Ultra settings the leads tighten and the Aorus Gaming 7 Takes the lead with a virtual tie for second between the Taichi and the Designare. At UHD, the spreads continue to tighten at high detail when compared to 1080p, and our worst-case Talos run gives the Designare the edge for average and max framerates.

With Metro Exodus on track for a PC release (pardon the pun), we can’t help but get excited every time we watch our bundled benchmark of Metro Last Light Redux run across our test display. Since we only factor in average framerates to our performance calculations, the Designare laags behind by roughly two frames at 1080p at both detail settings. At 4k, each of the tested motherboards shows practically identical results for all metrics, with slight variances observed in maximum frame rates.

Lastly, Ashes of the Singularity: Escalation continues to thrash the core counts of our 1950X, and shows a 2.9 percentage point loss at 1080p High and a 3.8 percentage loss for the Designare on the Crazy setting. As we observe in other games, increasing the resolution on this platform shows a tightening of the deltas across samples. But 4k High settings defeat the Designare here. On the bright side, the Designare nearly squeaks a 1 percentage point win against the Taichi at Crazy. But overall, each of these boards is more than adequate for gaming.

We have seen some close performance results in our previous AMD motherboard coverage, but this bunch of boards shows each board excelling in a few different areas. The Designare has a large lead compared to the competition in the synthetic realm, but lags in games which place it at the #2 spot in terms of overall combined performance. Interestingly enough, the Aorus Gaming 7 scored well with gaming and applications. The Taichi and Gaming Pro Carbon follow close behind, favoring gaming and applications respectively.

MORE: Best Motherboards

MORE: All Motherboard Content

Overclocking, Power, Thermal Benchmarks & Conclusion

Temperatures, Power and Efficiency

Ever since we transitioned to our Corsair 570X tempered glass enclosure, our temperature data has been straightforward. Using HWiNFO for sensor readings, the Tctl delta is a steady 27C, and overall all boards in the 570X are able to achieve CPU temperatures within 2 degrees of each other. More important for this chart are the regulator temperatures. Clearly the Taichi wins here, but it's interesting to see that the Designare is able to cool 5 degrees C better than the Aorus Gaming 7 with the same fan orientations. Be it fan speed differences, sample variation, or some other factor, the Designare’s design is well suited for overclocking and low-airflow cases.

Power consumption is always a difficult measurement to take and requires a lot of trial and error regarding background applications, operating voltages, and other environmental factors. For today’s comparisons, the Gigabyte X399 Designare is able to best its Aorus sibling in all categories and is comparable to the Taichi and Gaming Pro Carbon AC in all categories except complete system load. Again, this test is hard to control since we have to take into account Window’s ability to divert system scheduling to both power virus workloads as well as maintaining background processes. Regardless, the Designare performs well here, so let's see how she does with efficiency.

If we factor in the worst-case system load numbers from before, it’s tough to dethrone the ASRock X399 Taichi’s 20-watt lead compared to other boards. If we remove that metric from the chart, things become more visible, with the Aorus Gaming no longer being as bad from an efficiency perspective, while the MSI board slips due to its extra wattage under GPU load. However we spin the numbers, the Designare clearly performs better than its sibling and strikes a good balance compared to our other test boards.

Overclocking

Given its lineage and the success we had with the Gigabyte X399 Aorus Gaming 7, we expected that the Designare EX would perform well here. As always, increasing the multiplier to the 39.0 setting proves to be a simple task for this board with successful runs with both Prime95 and AIDA64.

Bumping to 3925MHz requires us to increase the loadline to the Aorus’ similar level with no additional voltage. With 4GHz still in our sights, logically we increased the multiplier to 39.5 and instantly achieved a black screen when loading Prime95. Increasing voltage appears to buy us some stability, but we hit our thermal threshold before we are able to gain our standard 8-hour acceptance time. With this defeat we continued to increase our multiplier to 39.75 and eventually 40.0 for 4GHz, but not even AIDA64 can run reliably at this point. Still, 3950MHz is a moderate overclock, and additional tuning might be able to squeeze more frequency or stability out of the system.

Armed with our Flare-X DDR4-3200 RAM kit, we engaged the default XMP profile for these DIMMs and had no issues running the AIDA64 memory stress test for 30 minutes. Seeing that our other boards hit DDR4-3333, we engaged that multiplier through the Gigabyte UEFI and saw mixed results across several attempts. Each reboot did work without the firmware re-trying its training routines, so it could just be that this particular motherboard sample couldn’t handle the extreme usage of AIDA64. As always, more tuning could probably get us to DDR4-3333 or higher, but as it stands we can engage XMP settings, which isn't always the case.

Value and Conclusion

If the performance data didn’t impress you and the thermal and power data didn’t help focus your sights on a board, then money might be the defining factor. At the MSRP, the Gigabyte X399 Designare EX can be yours for between $10 to $20 more than competing products. Given the cost of the complete system, that probably is acceptable given the added features bundled into the product. X399 is a robust and feature-rich environment, so the added aesthetics, VReg fan and dual network controllers nearly justify it from this reviewer’s eyes.

Now, let’s talk sale prices and speculation. At the time of writing, each of these boards can be between $30 and $80 cheaper depending on the retailer and rebates available. With that kind of variation, the Gigabyte X399 Aorus Gaming 7 becomes our value king. That make it unpalatable to purchase the Designare EX unless you're happen to spend $60 on this board's looks alone.

Don’t get us wrong, the Designare is a great board. In the stock configuration, it performs better than some competing products. If aesthetics are a key feature for your Threadripper build, this board wins hands down--unless of course you crave lots of included LEDs. If it could hit similar overclocking results compared to the Aorus Gaming 7, we’d likely hand it an award. But given current pricing and the pesky comparison data that came from our testing, we have to temper our enthusiasm for the X399 Designare EX a bit, no matter how pretty and feature-packed it is.

MORE: Best Motherboards

MORE: All Motherboard Content

Ryzen Threadripper 2 (2990WX and 2950X) Review: AMD Unleashes 32 Cores

palcorn@outlook.com (Paul Alcorn) — Mon, 13 Aug 2018 13:00:00 +0000

Introduction

AMD is building on the success of its high-end desktop Ryzen Threadripper processors with a handful of powerful new models known as the Threadripper 2000-Series or, unofficially, Threadripper 2.

The flagship Threadripper 2990WX is an $1800 beast armed with an incredible 32 Zen+-based cores and the ability to work on 64 threads concurrently. Threadripper 2970WX wields 24 cores and 48 threads, plus lower boost frequencies, but also pushes pricing down to $1300. Those are both quad-die configurations. AMD is also updating its dual-die line-up with the Threadripper 2950X, wielding 16 Zen+ cores for $900, and Threadripper 2920X, sporting 12 cores and the ability to work on 24 threads concurrently for $650. The 2950X, specifically, serves up impressive value across a diverse range of workloads. It may very well be the CPU to beat in today's high-end desktop market.

Even though AMD claims its top-end Threadripper 2990WX is up to 51% faster than Intel's $2000 Core i9-7980XE, there are some caveats enthusiasts need to be made aware of. For instance, the WX series' multi-chip module comprises four separate dies. Due to the constraints of AMD's existing Threadripper design, two of the 2990WX's dies aren't connected directly to main memory. That creates an architecture capable of incredible performance in heavily-threaded workloads that aren't sensitive to memory throughput, but less impressive results in bandwidth-hungry applications that don't scale well with extra cores. We found that the Threadripper 2990WX is mostly a niche product for professionals with specific requirements. Still, it sets a new high water mark for compute horsepower on the desktop.

Threadripper 2

Earlier this year, AMD retooled its mainstream Ryzen line-up with new Zen+ optimizations that include 12nm manufacturing, improved memory and cache latency, higher clock rates, and enhanced multi-core Precision Boost frequencies. The net effect of those changes carry over to the new Threadripper models.

And like the previous-gen Threadripper models, AMD offers improved frequency and voltage scaling by selecting the top 5% of Zeppelin dies for its halo product line. That should translate to lower voltage requirements at any given clock rate.

	Ryzen Threadripper 2990WX	Ryzen Threadripper 2950X
Socket	TR4	TR4
Cores / Threads	32 / 64	16 / 32
Base Frequency	3.0 GHz	3.5 GHz
Boost Frequency	4.2 GHz	4.4 GHz
Memory Speed	DDR4-2933 (Varies)	DDR4-2933 (Varies)
Memory Controller	Quad-Channel	Quad-Channel
Unlocked Multiplier	Yes	Yes
PCIe Lanes	64 (Four to the chipset)	64 (Four to the chipset)
Integrated Graphics	No	No
Cache (L2 / L3)	80MB	40MB
Architecture	Zen+	Zen+
Process	12nm LP GloFo	12nm LP GloFo
TDP	250W	180W

With today's introduction, AMD splits its Threadripper portfolio into the WX and X families. Similar to the company's previous-gen X products, the two new WX models are geared toward intense multitasking workloads, 3D rendering, media encoding, and cinema mastering. That makes them attractive to software developers, video/audio engineers, and content creators.

The 32C/64T Ryzen Threadripper 2990WX has a 3 GHz base frequency that stretches up to 4.2 GHz via AMD's XFR (eXtreme Frequency Range) algorithms. It does battle with Intel's $2000 Core i9-7980XE, boasting 18 Hyper-Threaded cores. AMD claims the 2990WX's single-threaded performance only lags the -7980XE by four percent. That's the closest the company has come to matching Intel in this important metric.

AMD also has a $1300 Threadripper 2970WX in the works that serves up 24 cores able to work on 48 threads simultaneously. However, it won't be available until October.

Each of the WX processors' four dies feature eight physical cores and 16MB of L3 cache. Threadripper 2990WX and 2970WX both enjoy the benefit of 64MB L3 cache. However, AMD disables two cores per die on the 2970WX to facilitate its 24-core configuration.

The WX family's more sophisticated layout results in a 250W thermal design power. Of course, that's significantly higher than the original Threadripper's 180W rating, meaning you need a high-end cooling solution to realize peak performance (particularly if you plan on overclocking). AMD ships all Threadripper CPUs with an Asetek bracket that provides partial coverage of the massive heat spreader using supported closed-loop liquid coolers. According to AMD, this partial coverage is fine for stock operation. But we found that full-coverage coolers work better. AMD also collaborated with Cooler Master to develop the Wraith Ripper heat sink/fan combo for its Socket TR4 interface. It's sold separately, though. The premium cooler comes with configurable RGB lighting and was designed to accommodate tall memory modules under its fin stack.

AMD uses Indium solder between its dies and heat spreader to improve thermal transfer. In contrast, Intel employs thermal grease. Intel also recommends liquid cooling on its Skylake-X processors, while, again, AMD contends that air cooling is ample.

	Cores /Threads	Base / Boost (GHz)	L3 Cache (MB)	PCIe 3.0	DRAM	TDP	MSRP	Price Per Core
TR 2990WX	32 / 64	3.0 / 4.2	64	64 (4 to PCH)	Quad DDR4-2933	250W	$1799	$56
TR 2970WX	24 / 48	3.0 / 3.2	64	64 (4 to PCH)	Quad DDR4-2933	250W	$1299	$54
Core i9-7980XE	18 / 36	2.6 / 4.4	24.75	44	Quad DDR4-2666	140W	$1999	$111
TR 2950X	16 / 32	3.5 / 4.4	32	64 (4 to PCH)	Quad DDR4-2933	180W	$899	$56
TR 1950X	16 / 32	3.4 / 4.4	64	64 (4 to PCH)	Quad DDR4-2667	180W	$750	$47
Core i9-7960X	16 / 32	2.8 / 4.4	22	44	Quad DDR4-2666	140W	$1699	$106
TR 2920X	12 / 24	3.5 / 4.3	32	64 (4 to PCH)	Quad DDR4-2933	180W	$649	$54
TR 1920X	12 / 24	3.5 / 4.2	64	64 (4 to PCH)	Quad DDR4-2667	180W	$399	$33
Core i9-7920X	12 /24	2.9 / 4.4	16.50	44	Quad DDR4-2666	140W	$1199	$100
Core i9-7900X	10 / 20	3.3 / 4.3	13.75	44	Quad DDR4-2666	140W	$999	$99
Core i7-8700K	6 / 12	3.7 / 4.7	12	16	Dual DDR4-2666	95W	$359	$60
Ryzen 7 2700X	8 / 16	3.7 / 4.3	16	16	Dual DDR4-2933	105W	$329	$41

Unlike AMD's first round of Threadripper chips, which were aimed at prosumers, this time the company is going after enthusiasts and gamers with its X series. The $900 Ryzen Threadripper 2950X weighs in with 16 cores and 32 threads, but features much higher clock rates than the WX models. The 2950X's 3.5 GHz base frequency and 4.4 GHz boost rate are slight steps up from Threadripper 1950X's 3.4/4.2 GHz specifications. Expect to find Threadripper 2950X in stock on August 31, 2018.

AMD plans to launch the $650 Threadripper 2920X in October. That 12-core, 24-thread CPU will offer a base frequency of 3.5 GHz and a maximum boost clock rate of 4.3 GHz. As with previous X-series models, the 2950X and 2920X utilize a pair of eight-core dies and two dummy packages that help with mechanical stability as you tighten down a thermal solution. The active dies expose 32MB of L3 cache, and abide the same 180W TDP rating as first-gen Threadripper processors.

All of the new Threadripper chips sport unlocked ratio multipliers for overclocking, along with 60 lanes of third-gen PCI Express (plus four lanes attached to the supporting chipset). All of that connectivity could come in handy for discrete graphics cards for rendering or compute, but they're also useful for high-performance storage and networking.

Threadripper CPUs feature independent dual-channel memory controllers located on two dies, which combine to provide quad-channel support with varying data transfer rates based upon your configuration. With today's introduction, AMD bumps its maximum specification up to DDR4-2933 from DDR4-2666. Today, the platform supports ECC memory and up to 256GB of capacity. However, it can accommodate up to 2TB as memory density increases.

DIMM Config	Memory Ranks	Official Supported Transfer Rate (MT/s)
4 of 4	Single	DDR4-2933
4 of 8		DDR4-2667
8 of 8		DDR4-2133
4 of 4	Dual	DDR4-2933
4 of 8		DDR4-2667
8 of 8		DDR4-1866

Precision Boost 2, XFR2, and Precision Boost Overdrive (PBO)

AMD's previous-gen Ryzen processors included features called Precision Boost, a dynamic voltage frequency scaling technology similar to Intel's Turbo Boost, and eXtended Frequency Range, which provided additional frequency uplift if your cooling solution had thermal headroom to spare. Those 1000-series CPUs only offered quad-core (X series) or all-core Precision Boost and XFR clock rates.

The Threadripper WX processor's highest boost frequency occurs on eight cores simultaneously, while the X-series chips boost on four cores. There remains headroom to exploit, though. Precision Boost 2, which debuted on the desktop with AMD's Raven Ridge processors, and XFR2 algorithms improve performance in threaded workloads by raising the frequency of any number of cores. Precision Boost 2 delivers up to 500 MHz-higher clocks during multi-core workloads, while XFR2 adds an additional 16% boost if your cooler is beefy enough.

AMD's Ryzen Threadripper 2990WX and 2950X launch marks the official introduction of Precision Boost Overdrive (PBO), a feature that rolled out quietly alongside second-gen Ryzen CPUs earlier in 2018. PBO is an automated overclocking feature that boosts performance to the limits of a motherboard and cooler's capabilities. Cruelly, using it voids Threadripper's three-year warranty. And AMD unfortunately has no equivalent to Intel's optional Performance Tuning Protection Plan.

The new Threadrippers already push the frequency/voltage curve's boundaries at stock settings, so manual tuning often result in worse single-threaded performance because the silicon can't accelerate as aggressively under an all-core overclock. Locking the processor to a static frequency also prohibits it from downshifting into lower clock rate to save power at idle. PBO addresses the issues with manual overclocking head-on. It dynamically overclocks the processor and communicates with the platform to modulate performance based on what the motherboard's power delivery subsystem can do. AMD doesn't share a list of specific multi-core Precision Boost 2, XFR2, and PBO bins because the opportunistic algorithms achieve different frequencies based on temperature, current, and load.

All of the 2000-series Threadripper processors are technically backward-compatible with existing X399 motherboards. With that said, power delivery is an important variable to consider, given the requirements of 32 cores versus 16. MSI and Gigabyte both announced new X399 motherboards with beefier power delivery subsystems, while Asus introduced an add-on kit for one of its existing motherboards that improves VRM cooling.

Part of the Threadripper package is, well, AMD's packaging. The original Threadripper box set a new bar for the industry (you can see our Threadripper 2 unboxing here), and the company stepped up its game with an even larger package this time around.

MORE: Best CPUs

MORE: Intel & AMD Processor Hierarchy - CPU Comparison with Benchmarks

MORE: All CPUs Content

Architecture, NUMA & Game Mode

It Starts With 12nm LP

AMD's Threadripper 2 processors are manufactured on GlobalFoundries' 12nm LP process technology. The ported-over design helps boost transistor performance, but does not affect die area or transistor density. As a result, the Zeppelin die's ~4.8 billion transistors and 213mm² area remain similar from first-gen Ryzen. The dual-die X-series models feature a total of 9.6 billion transistors and 426mm² of silicon, while the quad-die WX processors feature 19.2 billion transistors over 852mm².

Lower leakage current does enable 200 MHz-higher clock rates or an 80-120mV core voltage reduction at any given frequency compared to 14nm manufacturing. All told, AMD claims the 12nm design enables up to 11% less power consumption than 14nm-based Threadripper CPUs at the same clock rates, or up to 16% more performance at the same thermal design power. AMD also adds other nuanced refinements, like lower L1 (15%), L2 (9%), and L3 (8%) cache latencies, along with reduced memory latency (2%).

2990WX Architecture

Threadripper 2990WX borrows from AMD's EPYC server designs and comes with four active dies. The company fused off PCIe and memory control from two of the dies, creating silicon only useful for computing. Meanwhile, the other two I/O-enabled dies serve up two channels of DDR4 memory support and 32 lanes of PCIe 3.0 each.

Unfortunately, the compute dies suffer from increased latency on every request to main memory and PCIe-attached devices, as those requests always have to traverse the Infinity Fabric.

AMD added more Infinity Fabric channels to connect two more dies. Unfortunately, that has a tremendous impact on fabric bandwidth, which drops from 50 Gb/s on a 16-core Threadripper 2950X to 25 Gb/s in this implementation. And again, AMD measured performance with a 3200 MT/s data rate, meaning throughput at DDR4-2933 will be lower. Even with the benefits of tightly-controlled fabric scheduling magic, the combination of reduced bandwidth and 32 threads that must communicate over the fabric for I/O and memory requests has an impact on performance.

According to chip analyst David Schor at WikiChip, each request from the compute die requires interfacing with the Cache-Coherent Masters (CCM), which then interfaces with the CAKE (Coherent AMD socKet Extender) module that encodes the request and sends it to the remote I/O die. The remote CAKE module then decodes the request, fetches the requested data via the UMC (Unified Memory Controller), and then encodes the data and transmits it back to the compute die.

Increased traffic and reduced fabric throughput will have a tangible impact on memory-hungry applications, leading to sub-par performance scaling under some conditions. Although Threadripper 2990WX is clearly aimed at the semi-professional market, configurations hosting multiple GPUs may slow down due to increased fabric latency and reduced throughput to remote PCIe lanes. That'd also affect the performance of PCIe-based M.2 storage and LAN devices connected to remote dies.

MSI's MEG Creation motherboard diagram provides a nice summary of the split connectivity between dies. And be mindful of new population rules, such as inserting the first GPU into PCIe slot four, along with custom M.2 recommendations. You need to populate all four DRAM channels or follow dual-channel population rules in order to realize maximum performance, as performance drops sharply in some dual-channel configurations due to the distributed design.

AMD carves the Threadripper 2990WX into four NUMA domains that cannot be altered. As such, the processor does not have a local memory toggle for its Game Mode feature. Instead, the processor simply flips into "1/4" mode, which disables all but one die and effectively creates an 8C/16T CPU. Ryzen Master also has "1/2" and "Off" options that expose 16 cores and 32 threads, or 32 cores and 64 threads.

The company claims it could not enable the compute dies' memory and I/O controllers even partially without significantly overhauling the package's trace routing, requiring a new socket interface. AMD reps say they prioritized drop-in compatibility with the existing motherboard and cooler ecosystem, leading them to build Threadripper 2990WX the way it turned out.

AMD continues working with Microsoft to route threads to the die with direct-attached memory first, and then spill remaining threads over to the compute dies. Unfortunately, the scheduler currently treats all dies as equal, operating in Round Robin mode. As a result, even moderately-threaded applications can suffer at the hands of high memory latency and low throughput. This is further complicated by thread migration. According to AMD, Microsoft has not committed to a timeline for updating its scheduler.

The Zeppelin Building Block

The Zen architecture employs a four-core CCX (CPU Complex) building block. Each CCX has 8MB of L3 cache split into four slices; each core in the CCX accesses all L3 slices with the same average latency. Two CCXes come together to create an eight-core Zeppelin die, and they communicate with each other via AMD’s Infinity Fabric. The CCXes share the same dual-channel memory controller. This is basically two quad-core CPUs talking to each other over the Infinity Fabric pathway that also handles northbridge and PCIe traffic.

Although each core in a four-core CCX can access the local cache with the same average latency, trips to fetch data in adjacent CCXes incur a latency penalty. Communication between threads on cores located in disparate CCXes also suffers.

2950X Architecture & Game Mode

Threadripper 2950X mirrors the layout of AMD's first-gen Threadripper chips: two Zeppelin dies are connected via another layer of the Infinity Fabric. AMD flanks them with a pair of dummy dies that serve as non-functional fillers to ensure the heat spreader's structural integrity and consistent mating with the socket's pins.

Remember, each Zeppelin die has its own memory and PCIe controller. If a thread running on one core needs to access data resident in cache on another die, it has to traverse the fabric between those dies and incur significant latency. Naturally, the latency penalty between dies is higher than it is between CCXes in a single-die configuration. But AMD claims to have made some improvements there. The 2950X purportedly offers 64ns latency to near memory and 105ns to far memory, while the previous-gen 1950X had to wait 78ns and 133ns, respectively. As per usual, the speed of the Infinity Fabric is tied to the memory controller, so higher data rate settings are desirable. AMD measured Threadripper 2's fabric performance with a 3200 MT/s data rate, which means fabric latency at the recommended DDR4-2933 will be higher.

To combat the potential for performance regression as a result of its "go-wide" approach, AMD devised an interesting solution: it introduced a memory access switch that you can toggle via motherboard BIOS or the Ryzen Master software. The Local and Distributed settings flip between either NUMA (Non-Uniform Memory Access) or UMA (Uniform Memory Access), same as they did for AMD's first-gen Threadripper CPUs.

UMA (Distributed) is pretty simple; it allows both dies to access all of the attached memory. NUMA mode (Local) attempts to keep all data for the process executing on the die confined to its directly attached memory controller, establishing one NUMA node per die. The goal is to minimize requests to remote memory attached to the other die. NUMA works best if programs are designed specifically to utilize it. Even though most desktop PC software wasn't written with NUMA in mind, performance gains are still possible in non-NUMA applications.

AMD also allows you to disable cores in Legacy Compatibility mode, which disables one die via a Windows command. This allows some programs that won't function with 32 threads to execute properly, and it also eliminates cross-die communication. The system can still access I/O connected to the second die, though, so you don't lose any associated memory or attached peripherals.

A set of toggles generally offers the best performance in games and applications by combining these settings optimally. Game Mode disables one die with the Legacy Compatibility mode, and then switches the 2950X into Local memory mode, effectively creating an 8C/16T CPU. Creator mode uses the Distributed memory setting and disables Legacy Compatibility, providing access to Threadripper 2950X's full armament of 16 cores and 32 threads for demanding workloads.

MORE: Best CPUs

MORE: All CPUs Content

AMD Ryzen Threadripper 2990WX

AMD Ryzen Threadripper 2950X

Ryzen Master, Motherboards & Test Setup

Ryzen Master

AMD's second-gen Threadripper processors communicate with the platform to modulate performance based on what the motherboard's power delivery subsystem can do, which is the key enabler for Precision Boost Overdrive.

The processor monitors Package Power Tracking (PPT) and Thermal Design Current (TDC) variables, measuring available margin to the motherboard's maximum power output and current, respectively. Electrical Design Current (EDC) also indicates the maximum current possible from the VRMs during peak/transient conditions. A control loop feeds real-time telemetry data back to the Infinity Fabric, which then allows the processor to affect performance based on thermal and power conditions dynamically. AMD exposes some of these monitoring features with its updated Ryzen Master overclocking software.

Each motherboard vendor defines its own maximum amperage for PBO based on the power delivery subsystem's capabilities. As you can see in the Ryzen Master screenshot below, MSI has its own custom variables assigned for EDC, TDC, and PPT limits in the MEG X399 Creation motherboard. Due to different settings on different motherboards, PBO performance will vary.

You can also use Ryzen Master to switch the X-series processors into Game or Creator Mode, and toggle the WX-series chips into various legacy compatibility modes. AMD's software now supports per-CCX overclocking as well, and includes a built-in stress test. Again, the three-year warranty does not cover damage caused by overclocking, so exercise caution.

X399 Motherboards

Threadripper 2 is compatible with the X399 chipset, and AMD guarantees that all existing motherboards support stock operation. You do need an updated BIOS. However, all X399-based motherboards have an out-of-band BIOS update mechanism, so you won't need one of AMD's Boot Kits. Here's the rub: while all X399 motherboards support PBO (it's baked into the CPU), many existing platforms have limited power delivery subsystems that weren't designed to accommodate the 32C/64T 2990WX. Ultimately, they'll limit overclocking headroom. If you plan to overclock a second-gen Threadripper or enable PBO, it would be wise to verify the capacity of your motherboard's power circuitry first.

If you're building a first Threadripper-based PC, MSI has a new MEG X399 Creation. Moreover, Gigabyte recently introduced its X399 Aorus Xtreme. Both boards feature capable voltage regulation subsystems and robust cooling. Asus also introduced a cooling kit for its existing ROG X399 Zenith Extreme, which should bring that board up to par with the WX family's greater power consumption.

To recap, the X399 chipset supports two USB 3.1 Gen2 and six USB 3.1 Gen1 ports, along with six USB 2.0 connections. Two PCIe 3.0 lanes allow motherboard vendors to add more storage connectivity (four SATA or two SATA Express), and the eight general-purpose PCIe 2.0 lanes accommodate other controllers, such as Ethernet or WLAN/Bluetooth. Eight SATA ports round out the chipset's connectivity options, and you can leverage several RAID configurations with the attached SATA devices.

The Threadripper processor provides an additional eight USB 3.1 Gen1 ports and four SATA connections (hardware RAID supported). The 60 remaining PCIe lanes support up to seven PCIe devices. Threadripper 2 CPUs also benefit from an improved SensMI suite, including StorMI Technology. That's a software-based tiering solution able to meld the low price and high capacity of hard drives with the speed of SSDs, 3D XPoint (including Intel's Optane parts), or even up to 2GB of RAM. For more information, check out AMD and Enmotus Expand FuzeDrive Offerings.

Comparison Products

Intel Core i9-7960X

Intel Core i9-7980XE

AMD Ryzen 7 2700X

Test Setup

You naturally need a capable power supply to support Ryzen Threadripper 2990WX. After a moderate 3.8 GHz overclock, we observed as much as 38A of current draw during a threaded Cinebench test. Even in its stock configuration, the CPU can pull more than 20A in that benchmark. And if you enable PBO, current draw jumps to a similar level as if you were overclocking manually.

We tested the Threadripper 2 models with MSI's MEG X399 Creation motherboard. Due to cooling and power delivery constraints, we ran through our full test suite at stock settings and with PBO activated, rather than using an all-core overclock. Our PBO-enabled configurations did benefit from higher memory transfer rates, as detailed in the table below.

Test System & Configuration
Hardware	Germany AMD Socket AM4 (400-Series)AMD Ryzen 7 and Ryzen 5 MSI X470 Gaming M7 AC 2x 8GB G.Skill FlareX DDR4-3200 @ DDR4-2667, DDR4-3466AMD Socket SP3 (TR4)Threadripper 2MSI MEG X399 Creation 4x 8GB G.Skill FlareX DDR4-3200 RGBIntel LGA 1151 (Z370): Intel Core i7-8700K MSI Z370 Gaming Pro Carbon AC 2x 8GB G.Skill FlareX DDR4-3200 @ DDR4-2667, DDR4-3466Intel LGA 2066 Intel Core i7, Core i9 MSI X299 Gaming Pro Carbon AC 4x 8GB G.Skill FlareX DDR4-3200 @ DDR4-2666All SystemsGeForce GTX 1080 Founders Edition (Gaming) Nvidia Quadro P6000 (Workstation)1x 1TB Toshiba OCZ RD400 (M.2, System SSD) 4x 1TB Crucial MX300 (Storage, Images)be quiet! Dark Power Pro 11, 850W Windows 10 Pro (All Updates)U.S. AMD Socket SP3 (TR4)Threadripper Gen 1 & 2MSI MEG X399 Creation 4x 8GB G.Skill FlareX DDR4-3200 @ DDR4-2933, DDR4-3200Intel LGA 2066Intel Core i9-7960X, -7980XE, -7900XMSI X299 Gaming Pro Carbon AC4x 8GB G.Skill FlareX DDR4-3200 @ DDR4-2666, DDR4-3200AMD Socket AM4 (400-Series)AMD Ryzen 7 2700X MSI X470 Gaming M7 AC2x 8GB G.Skill FlareX DDR4-3200 @ DDR4-2933Intel LGA 1151 (Z370)Intel Core i7-8086K, Core i7-8700K, Core i5-8600K, Core i5-8400, Core i7-8700MSI Z370 Gaming Pro Carbon AC2x 8GB G.Skill FlareX DDR4-3200 @ DDR4-2667All EVGA GeForce GTX 1080 FE 1TB Samsung PM863SilverStone ST1500-TI, 1500WWindows 10 Pro (All Updates)
Cooling	GermanyAMD Wraith RipperAlphacool Ice Block XPXEnermax LiqTech 240 TR4Thermal Grizzly KryonautU.S.Wraith RipperCorsair H115iEnermax Liqtech 240 TR4 II
Power Consumption Measurement	Contact-free DC Measurement at PCIe Slot (Using a Riser Card) Contact-free DC Measurement at External Auxiliary Power Supply Cable Direct Voltage Measurement at Power Supply 2x Rohde & Schwarz HMO 3054, 500 MHz Digital Multi-Channel Oscilloscope with Storage Function4x Rohde & Schwarz HZO50 Current Probe (1mA - 30A, 100 kHz, DC) 4x Rohde & Schwarz HZ355 (10:1 Probes, 500 MHz) 1x Rohde & Schwarz HMC 8012 Digital Multimeter with Storage Function
Thermal Measurement	1x Optris PI640 80 Hz Infrared Camera + PI Connect Real-Time Infrared Monitoring and Recording
Acoustic Measurement	NTI Audio M2211 (with Calibration File, Low Cut at 50Hz) Steinberg UR12 (with Phantom Power for Microphones)Creative X7, Smaart v.7 Custom-Made Proprietary Measurement Chamber, 3.5 x 1.8 x 2.2m (L x D x H) Perpendicular to Center of Noise Source(s), Measurement Distance of 50cm Noise Level in dB(A) (Slow), Real-time Frequency Analyzer (RTA) Graphical Frequency Spectrum of Noise

MORE: Best CPUs

MORE: All CPUs Content

AMD Ryzen Threadripper 2990WX

AMD Ryzen Threadripper 2950X

VRMark, 3DMark & AotS: Escalation

Test Notes

AMD designed Threadripper 2990WX for prosumer-class applications. Unlike the previous-gen Threadripper models, its WX-series models come with a Game Mode preset in the Ryzen Master software that disables three of the four available dies (1/4). AMD tells us this offers the best average performance in a wide range of titles. But the company also provides toggles that allow experimentation with two- and four-die configurations.

Unlike the first-gen Ryzen Threadripper 1950X, AMD aims its 2950X at enthusiasts and gamers. For this review, we tested Threadripper 2950X using AMD's Game Mode setting. But in our dedicated review of that chip, we'll go into more depth on the available combinations of settings and their impact on performance.

We tested across our gaming suite using a 1920x1080 resolution, minimizing graphics bottlenecks. Of course, as you step up to 2560x1440 or 3840x2160, the difference between processors shrinks. Just bear in mind that, beyond the average frame rates we report, Ryzen Threadripper 2990WX and 2950X are also well-suited to gaming while multi-tasking and streaming due to their high core counts.

VRMark, 3DMark

We aren't big fans of using synthetic benchmarks to measure game performance, but 3DMark's DX11 and DX12 CPU tests provide useful insight into the amount of horsepower available to game engines.

Moreover, UL's VRMark test lets you gauge your system's suitability for use with the HTC Vive or Oculus Rift, even if you don't currently own an HMD. UL defines a passing score as anything above 109 FPS.

Tests that are sensitive to clock rate and IPC throughput, such as VRMark, were a challenge for AMD's first-gen Threadripper processors. But we saw a big improvement from Threadripper 2950X compared to the previous-gen 1950X, which was expected due to the more aggressive multi-core turbo bins.

The 2990WX's Game Mode reduces overall core count, but it also keeps bandwidth-starved cores from hurting performance. Nevertheless, Ryzen Threadripper 2990WX fell to the bottom of our chart due to lower per-core performance. Enabling PBO helped push it up to the middle of our test field.

3DMark typically scales well with higher core/thread counts. But the Threadripper processors, including the 32C/64T 2990WX, lagged Intel's line-up. The 2950X did enjoy a nice speed-up compared to AMD's older Threadripper 1950X. However, the 2990WX was hobbled by its Game Mode setting that turned it into an 8C/16C CPU. Both Threadripper 2 models realized solid gains from enabling Precision Boost Overdrive.

Ashes of the Singularity: Escalation

Ashes of the Singularity: Escalation is a computationally intense title that scales well with thread count.

Ryzen Threadripper 2990WX lagged the rest of our test pool at its stock settings, but matched an overclocked 1950X with PBO enabled.

Meanwhile, the 2950X scored another solid victory against AMD's previous-gen 1950X. But neither model comes close to matching Intel's highest-end processors.

This is one of the best examples of a game that scales well with host processing resources. However, the fact that Ryzen 7 2700X outperformed most of the test pool at a significantly lower price is telling. It's best to stick with mainstream desktop CPUs if gaming is your primary goal.

MORE: Best CPUs

MORE: All CPUs Content

AMD Ryzen Threadripper 2990WX

AMD Ryzen Threadripper 2950X

Civilization VI Graphics & AI, Dawn of War III

Civilization VI AI Test

Civilization VI's AI test measures CPU performance in a turn-based strategy game and tends to favor per-core performance.

Intel's high-end desktop roster dominated the Civilization VI AI benchmark.

Again, Threadripper 2950X fared well compared to previous-gen Threadripper models.

Although Ryzen Threadripper 2990WX landed behind its overclocked predecessors, automated tuning yielded a respectable performance boost.

This test continues to favor Intel architectures, and it's noteworthy that we don't have overclocked Core i7 and Ryzen 7 models in our chart to keep analysis simpler. But those CPUs offer the best value when gaming is your priority.

Civilization VI Graphics Test

Ryzen Threadripper 2990WX beat the stock 1950X and 1920X. And after enabling PBO, it also beat those two models overclocked as high as we dared take them.

Warhammer 40,000: Dawn of War III

Ryzen Threadripper 2950X and 2990WX were much more competitive in this game once we enabled their PBO feature.

MORE: Best CPUs

MORE: All CPUs Content

AMD Ryzen Threadripper 2990WX

AMD Ryzen Threadripper 2950X

Far Cry 5, GTA: V & Hitman

Far Cry 5

Ryzen Threadripper 2990WX dragged along the bottom of our chart at its stock settings. Enabling PBO helped quite a bit, though.

Threadripper 2950X was just a bit faster. That's particularly good news for gamers, given a much lower price tag.

Grand Theft Auto V

Grand Theft Auto V favors Intel architectures and, more generally, multi-core designs with high clock rates.

Threadripper 2950X beat the Core i9-7900X after we enabled Precision Boost Overdrive. Then again, Intel's Core i7-8700K and AMD's Ryzen 7 2700X reminded us why we don't recommend installing high-end desktop CPUs in most gaming PCs.

Hitman

Our Hitman benchmark was rendered almost useless by a patch that imposed a 90 FPS performance cap. A subsequent update restored our Hitman test to its prior glory.

Hitman responds to core count and clock rate, so it wasn't a surprise to see Intel's overclocked Core i9-7960X perform best in our benchmark. Otherwise, the results landed where we expected them to.

MORE: Best CPUs

MORE: All CPUs Content

AMD Ryzen Threadripper 2990WX

AMD Ryzen Threadripper 2950X

Shadow Of War & Project CARS 2

Middle-earth: Shadow Of War

Both second-gen Threadripper processors were highly competitive in this game (though Ryzen Threadripper 2950X unsurprisingly offered better performance).

Project CARS 2

Although Project CARS 2 is purportedly optimized for threading, clock rates obviously affect frame rates. As a result, Intel's Core i7-8700K kept pace with much more expensive alternatives.

We observed a repeatable spike in our Threadripper 2990WX benchmark results. However, it had little impact on overall performance and registered as only a slight dip in the 99.9th percentile frame time.

MORE: Best CPUs

MORE: All CPUs Content

AMD Ryzen Threadripper 2990WX

AMD Ryzen Threadripper 2950X

Office & Productivity

Adobe Creative Cloud

Even though this suite has a few parallelized workloads, the final score is heavily influenced by the lightly-threaded tasks common in most desktop applications. The Threadripper 2950X outperforms its predecessor handily at stock settings, but the tuned Threadripper 1920X’s 4.0 GHz clock speed steals the show. The 2990WX languishes at the bottom of the overall score chart at stock settings, largely due to its lower frequencies, but tuning boosts it into contention with the Core i9-7980XE.

The Threadripper 2950X generally offers a more balanced profile than the 2990WX in most tests, but the 2990WX’s strength in the Indesign and Photoshop Heavy tests contribute to its overall lead. The tuned 1920X regularly pops up over the 2950X, but we verified the results through extensive retests.

Web Browser

The Krakken suite tests JavaScript performance using several workloads, including audio, imaging, and cryptography. Like most web browser workloads, single threaded performance reigns supreme here. As such, it’s not surprising to find the Core i7-8700K and the Ryzen 7 2700X
at the top of the pile.

The MotionMark benchmarks, which emphasize graphics (rather than JavaScript), are also exceedingly sensitive to CPU clock rates. Intel’s processors take the uncontested lead. The same story plays out in the WebXPRT benchmark.

Productivity

The application start-up metric measures load time snappiness in word processors, GIMP, and Web browsers under warm- and cold-start conditions. Other platform-level considerations affect this test as well, including the storage subsystem. This benchmark remains firmly in Intel’s favor, and once again we notice the Threadripper 1920X’s 4.0 GHz clock speed boosting it above its newer counterparts.

Our video conferencing suite measures performance in single- and multi-user applications that utilize the Windows Media Foundation for playback and encoding. It also performs facial detection to model real-world usage. Ryzen 7 2700X takes the lead while the 2990WX continues to lag in tests that aren’t heavily parallelized, but tuning brightens the picture somewhat.

The photo editing benchmark measures performance with Futuremark's binaries using the ImageMagick library. Common photo processing workloads also tend to be parallelized and the 2990WX rises to the occasion. The processor easily beats out the rest of the test pool, but
considering its 32 cores, the slight advantage over the tuned 16C Threadripper 1950X reminds us that AMD’s new flagship's performance doesn’t always scale well.

MORE: Best CPUs

MORE: All CPUs Content

AMD Ryzen Threadripper 2990WX

AMD Ryzen Threadripper 2950X

Rendering, Encoding & Compression

Rendering

Threaded workloads remain an uncontested strength of AMD's Zen-based processors and their hefty core counts. Many of these workloads stress the memory subsystem, which reduces the advantage of the Threadripper 2990WX’s hefty core count due to accesses from the remote memory controllers.

The Ryzen line-up dominates the multi-core Cinebench and POV-Ray tests, but the 2990WX only provides a 35% speed over the 2950X in the POV-Ray benchmark. In light of its 100% increase in cores, that doesn’t represent the best scaling possible. The 2990WX provides a 66% performance improvement in the Cinebench score.

Corona also scales well as the Threadripper 2990WX unleashes the full might of its 64 threads. Intel’s processors still hold the per-core advantage in the single core POV-Ray and Cinebench tests, but its easy to see that AMD’s multi-die design can help offset that advantage with extra cores in threaded workloads.

Encoding & Compression

Our threaded compression and decompression tests work directly from system memory, removing storage throughput from the equation. This workload benefits heavily from threading, but either memory throughput or poor software scaling is holding the 29990WX back from realizing
its potential in the compression test. Threadripper 2950X, which comes with two die outfitted with directly-attached memory controllers, offers a nice boost over its predecessor at stock settings. The Core i9-7980XE reigns supreme in this test, which could be due to its single monolithic die and mesh architecture that facilitates much higher bandwidth to the cores. In contrast, the decompression test highlights the 2990WX’s devastating performance when the cores are properly fed.

y-cruncher, a single- and multi-threaded program that computes pi using AVX instructions, is a great test to measure Threadripper 2’s AVX performance. Intel’s Core i9 employs two 256-bit AVX FMA units per core that operate in parallel, whereas Ryzen's Zen architecture divides 256-bit AVX operations across two FMA units per core. That means the 2990WX’s two extra die bring much more AVX performance to the table, provided you can feed the data-hungry AVX operations. The 2990WX nears the top of the chart, but scaling is sub-optimal compared to the 2950X, which has half the AVX units. This could again fall back on poor memory subsystem performance, which hinders some applications.

That same story plays out in Handbrake. AMD said during its briefings that Handbrake is cache dependent and doesn’t utilize cores fully, thus yielding a smaller performance boost than expected. The Handbrake x265 test, which uses a heavier distribution of AVX instructions than the H264 test, yields a minimal performance advantage for the 2990WX compared to the 2950X. We think that memory access plays a larger role in this result than sub-optimal software scaling.

MORE: Best CPUs

MORE: All CPUs Content

AMD Ryzen Threadripper 2990WX

AMD Ryzen Threadripper 2950X

Workstation - Graphics

Anyone interested in using a Threadripper processor in a workstation should carefully consider the applications they're running. Over and over, we're reminded that not all tasks can be effectively parallelized, and many workloads are only optimized for four to eight cores.

AutoCAD is a prime example. If you work in 2D draft mode, it's rare to see more than two cores utilized. There, IPC throughput wins over core count. Ryzen Threadripper 2950X can at least push out in front of Ryzen 7 2700X. However, Ryzen Threadripper 2990WX looks out of place, given a much higher price tag.

Our results were similar in the Cinebench graphics benchmark, which combines host processing and 3D workloads. Faced with a slightly more demanding benchmark, Ryzen Threadripper 2990WX with PBO enabled hit a high enough clock rate to match Threadripper 2950X.

In the end, though, Intel's line-up fared far better.

SolidWorks consists of several components and packages that utilize host processing resources differently. In the graphics-bound composite, which is not optimized for threading, AMD's portfolio was generally routed.

AMD's extra cores were a little more effective in the Creo graphics composite test. But again, Intel dominated absolutely.

Catia is similar to the Cinebench graphics test, so our results weren't particularly surprising. Ryzen Threadripper 2950X took a wafer-thin lead over the Ryzen 7 2700X, demonstrating that its extra cores don't trip over themselves in lightly-threaded workloads.

This Blender workload doesn't utilize lots of execution cores, but at least it put AMD's newest Threadripper CPUs ahead of Ryzen 7 2700X.

MORE: Best CPUs

MORE: All CPUs Content

AMD Ryzen Threadripper 2990WX

AMD Ryzen Threadripper 2950X

Workstation - Compute

Whereas rendering tasks are typically graphics-heavy, favoring CPUs with high clock rates, compute workloads are more up Threadripper's alley. The new 2950X is definitely a step up from AMD's previous-generation Threadripper 1950X. Meanwhile, Ryzen Threadripper 2990WX doesn't always scale well, making it difficult to justify that premium price.

In SolidWorks, for example, Threadripper CPUs failed to impress since our benchmark didn't scale well across available cores. It's also no secret that many applications were compiled and optimized for Intel CPUs, putting AMD at an inherent disadvantage.

Once we fire up a benchmark that's able to utilize all available processing resources, AMD's Ryzen Threadripper chips take off. The 2990WX carved out a seemingly insurmountable lead, while the 2950X with PBO enabled nudged past Intel's Core i9-7980XE.

The 3ds Max SPECapc workload is now dominated by AMD, which uses physical cores to outmaneuver Intel's mix of real and logical resources. The Ryzen 7 2700X did score a first-place finish thanks to its higher clock rate, though.

Flip over to the pure rendering score and Threadripper 2950X with PBO enabled jumps into first place. There, Ryzen 7 2700X is easily outperformed by multiple Threadripper configurations.

The LuxRender workload from SPECwpc is well-optimized for threading, allowing AMD's Ryzen Threadripper 2990WX to truly shine. Every other model is left in the dust.

HandBrake does scale across multiple cores, but it tops out before the 32C/64T Threadripper 2990WX really takes off. As a result, the flagship Threadripper model loses to AMD's 2950X with PBO enabled.

7-Zip is similar in that our workload doesn't scale perfectly according to core/thread count. In fact, this benchmark is influenced more heavily by clock rate. Eight threads are usually sufficient, allowing Core i7-8700K to claim a first-place finish.

MORE: Best CPUs

MORE: All CPUs Content

AMD Ryzen Threadripper 2990WX

AMD Ryzen Threadripper 2950X

Cooling, Clock Rates & Power Limits

At first, we were amused by AMD's suggestion that these new Ryzen Threadripper CPUs could be cooled effectively with heat sinks and fans. But wouldn't it be something if the company's representatives were correct and we could cover the 2990WX with a beefy-enough sink to keep it running at full speed under load, without throttling issues? After all, we've already seen Intel's flagship stagger under the limitations of thermal grease between its heat spreader and die. Might AMD's use of Indium solder pay off during such an experiment?

Air Cooling And Its Limits

As it turns out, you really can cool the 180W Ryzen Threadripper 2950X and the 250W Ryzen Threadripper 2990WX using a heat sink and fan. It's even possible to leave PBO turned on with Threadripper 2950X, though noise starts becoming an issue under loads where power consumption peaks around 250W.

Although the Cooler Master-supplied heat sink is plenty capable, we have to imagine anyone willing to air-cool such an expensive CPU probably wants a potent Noctua sink with a quieter fan instead. What annoyed us most about AMD's provided cooler was the sleeved cable coming out the side facing your first PCIe add-in card. If you put a thick, high-end board in that first slot, then the cooler's USB port is completely covered and no longer usable. Additionally, a graphics card with a backplate touches the sink, which shouldn't happen.

At this point, we're switching to water cooling, and once the thermal limit of these processors is reached, we'll use an even more powerful chiller system.

Of course, if you're hoping to realize the maximum benefit of XFR2 and PBO, it's important to use the right cooler, and to back it with a high-end power supply. After taking the 2990WX to its limits, we hooked up the chiller and measured more than 42A of current on the motherboard's EPS connectors. Under LN2 cooling, current draw exceeded 50A and we had to circumvent the motherboard's 500W limit.

Both models appear to use a new temperature detection and reporting mechanism compared to the previous generation. Although we do see a 27°C-higher Tctl value, it only exists for backward compatibility on functions like the fan control circuit. Otherwise, both Threadripper models have an upper Tdie limit of 68°C, which would yield a Tctl value of 95°C. Tctl is therefore no longer a measured variable.

The curves for Ryzen Threadripper 2990WX appear in the graph below. With the heat sink/fan and closed-loop liquid cooler, this 32-core beast doesn't get very far under a taxing load.

It's easy to applaud Ryzen Threadripper 2950X's performance, since you don't have to put as much effort into cooling it. Whereas the 2990WX quickly ran out of steam under air or a closed-loop liquid-cooling setup, this model behaves well even if you top it with an all-in-one setup.

While a tuned Ryzen Threadripper 2950X can be cooled at just over 250W using a heat sink and fan (and still operate properly), you really want a liquid-cooler of some sort to get the best performance from it. But without PBO enabled, air cooling is effective enough, even up to 180W.

When it comes to Ryzen Threadripper 2990WX without PBO turned on, it's possible to hit that same 250W threshold and achieve sufficient cooling performance with a heat sink and fan. If you want to overclock much further, invest in a more capable thermal solution.

MORE: Best CPUs

MORE: All CPUs Content

AMD Ryzen Threadripper 2990WX

AMD Ryzen Threadripper 2950X

Power Consumption

It appears that AMD made a conscious effort to minimize idle power consumption compared to previous-generation Threadripper CPUs. This is most impressive from Ryzen Threadripper 2990WX, which hosts four active dies. Clearly, that model ducks in well under the level you'd expect by multiplying Ryzen 7 2700X's idle power consumption by four.

Threadripper 2950X fared especially well during our longer CAD run, which predominantly only uses up to four cores.

The 2990WX, on the other hand, sucks down a lot more power without a commensurate performance improvement.

Our gaming workload reflects big gains from AMD's second-gen Threadripper CPUs compared to their predecessors. Even turning PBO on for some extra performance doesn't kill the power story. Both new models offer significant efficiency improvements.

Full load represents a worst-case scenario for any flagship-class CPU.

Both new Ryzen Threadripper models employ Indium solder between their dies and heat spreader, whereas Intel sticks with thermal grease. We delidded CPUs from both families in order to measure overclocked power consumption without thermal throttling ruining our readings. Otherwise, we would have hit a ceiling at around 300W with Intel's grease under the hood.

It's apparent that motherboards impose AMD's specifications as hard limits: Threadripper 2950X, 1950X, and 1920X all top out at 180W without PBO enabled, while 2990WX peaks at 250W and not a watt more.

Power management is therefore the real highlight of today's launch, especially since the influence of cooling was perfectly implemented.

Even during our stress test with PBO enabled, Ryzen Threadripper 2990WX's thermal battle ends exactly at 500W. To top this, you need to manipulate your motherboard's limits and start messing with LN2. Then it's possible to top out just under 600W using a static 4.1 GHz.

Really, Threadripper 2950X represents the sensible upper limit for daily use. It also happens to be economically viable at a price point around $900. Compared to the previous-gen Threadripper flagship, this new 16C/32T model is a big improvement. Ryzen Threadripper 2990WX, on the other hand, just doesn't impress as much.

MORE: Best CPUs

MORE: All CPUs Content

AMD Ryzen Threadripper 2990WX

AMD Ryzen Threadripper 2950X

Final Analysis

Ryzen Threadripper 2990WX continues AMD's assault on Intel's market dominance with 32 cores and the ability to work on 64 threads concurrently. But AMD's new flagship has its issues, too. There's no denying the allure of such a powerful processor. But as we've seen from other high-core-count CPUs, power delivery and thermals can conspire to hinder performance.

While drop-in compatibility with existing X399 motherboards is a big selling point favoring AMD, not every board's power delivery subsystem is up to the job of facilitating maximum performance from the 2990WX (particularly if you plan to overclock). You really need a high-end motherboard, a high-end power supply, and high-end cooling to extract the utmost performance from your investment.

Ryzen Threadripper 2990WX is nothing short of phenomenal in workloads that can utilize its arsenal of execution cores. Rendering is a good example. But it doesn't scale well in other applications. The distributed architecture leaves half of the processor's compute resources stranded from the memory and I/O controllers, which can drastically reduce performance in applications sensitive to bandwidth or PCIe traffic. Unless you have a very specific workload that can't get enough parallelism, you're better off with Threadripper 2950X.

Threadripper 2950X is much more appealing to the enthusiast audience. Its dual-die design once seemed exotic, but now proves nimble in a wide variety of applications. Moreover, higher Precision Boost frequencies make the 2950X competitive against well-established Intel Core processors in lightly-threaded metrics.

Precision Boost Overdrive offers an easy path to overclocking both new Threadrippers if your supporting hardware can handle the stress. Just be aware that any gains you might see vary based on your system's capabilities. We're sure we can tease even more performance out via manual overclocking when we circle back for the dedicated review, but the automated overclocking feature is plenty capable for most users, and you retain the benefit of a beastly 4.4 GHz quad-core boost frequency. The Threadripper 2950X brings a lot more power for a lower price than the first-gen Threadripper did at launch, but we wouldn't recommend a direct upgrade from the 1950X. If you're looking to upgrade from an older CPU to an all-around crowd pleaser, Threadripper 2950X does not disappoint.

Considering the raw horsepower on offer, AMD's pricing is extremely competitive. Ryzen Threadripper 2950X kept pace with the $1700 Core i9-7960X in many of our tests, but sells for almost half of its price. Both Intel and AMD HEDT platforms are expensive, but X399 motherboards are particularly pricey, which you'll have to consider when weighing your options. Populating all four memory channels will also be expensive in these trying times, but that extra cost applies to both high end platforms.

If you're after the utmost in threaded performance for the dollar, the Ryzen Threadripper 2 series delivers. AMD clearly takes the lead with the most raw computing power on the desktop PC market, and at ultra-competitive pricing. We can't wait to see Intel's response.

MORE: Best CPUs

MORE: All CPUs Content

AMD Officially Opens Threadripper 1950X Exchange Program For Winners Of Intel's Core i7-8086K

palcorn@outlook.com (Paul Alcorn) — Mon, 25 Jun 2018 19:24:03 +0000

AMD's announcement last week that it would swap 40 Core i7-8086K's won from Intel's sweepstakes with a much beefier Threadripper 1950X CPU was an unexpected escalation in the long-running feud between the two companies. Now AMD has followed through with its plan and officially opened its replacement program at www.amd.com/thenext40. Per AMD's website:

AMD's "contest" is open to individuals that have won the Intel Core i7-8086K, a special edition processor that commemorates the 40th anniversary of the 8086, through Intel's sweepstakes.

Entrants are required to forward their prize notification from Intel to TR.Exchange@amd.com for verification. The first 40 verified winners can then exchange their new unopened processor for a Ryzen Threadripper 1950X. You can currently buy the Core i7-8086K for ~$425 at retail outlets, but those processors don't qualify for AMD's exchange program.

The $799 Threadripper 1950X wields 16-cores and 32-threads, which easily outweighs the i7-8086K's six-cores and twelve-threads in threaded heavily-workloads, such as those found in many productivity applications. But it is also noteworthy that "winners" of AMD's competing sweepstakes will have to pony up for a much more expensive X399 motherboard with the TR4 socket, which currently retails for more than $300, instead of Intel's less-expensive 300-series motherboards. Regardless, those who do swap their Intel Core silicon for an AMD Threadripper chip will gain ten cores and quad-channel memory, not to mention quite a bit of resale value.

The Core i7-8086K's superior single-threaded performance provides leading performance in most games, so we assume most gamers won't be interested in the exchange program.

Due to laws and regulations, AMD's replacement program is only open to legal residents of the United States that are over the age of 18. AMD requires winners to ship their unopened processor to the company, which will then ship out the Threadripper 1950X within four to six weeks. That's quite the wait for eager winners. AMD also states that "once you make the exchange by sending AMD your Competitor Sweepstakes prize, we’re not going to send it back if you change your mind."

The burning question in everyone's mind is what AMD will do with the exchanged Core i7-8086K processors. It's possible that AMD will use pictures of the 40th-anniversary chips for marketing purposes, but the fate of the exchanged processors is anyone's guess.

AMD Trolls Intel: Offers 16-Core Chip to Winners of Six-Core 8086K (Updated With Intel Response)

palcorn@outlook.com (Paul Alcorn) — Mon, 18 Jun 2018 23:20:00 +0000

EDIT: Intel responded to AMD's program. We've included the company's comments at the bottom of the article.

AMD's feud with Intel took an interesting turn today as the company announced that it would swap 40 Core i7-8086K's won from Intel's sweepstakes with a much beefier Threadripper 1950X CPU.

At Computex 2018, Intel officially announced it was releasing the Core i7-8086K, a special edition processor that commemorates the 40th anniversary of the 8086, which debuted as the first x86 processor on June 8, 1978.

As part of the special-edition release, Intel opened up a sweepstakes to give away 8,086 of the six-core 12-thread processors. Intel also made the processors available at retail, and though the company doesn't have an official MSRP, you can find the chips at several retailers for ~$425.

Now AMD is offering to replace 40 of the winners' chips with its own 16-core 32-thread $799 Threadripper processors, thus throwing a marketing wrench into Intel's 40th-anniversary celebration. We can't say it better than AMD's own snarky announcement:

Our competitor recently kicked off a sweepstakes to celebrate the first 40 years of the x86 processor by giving away 8,086 limited edition processors.We appreciate the advancements they’ve helped drive with the x86 architecture over the last four decades. But, we’re ready to take it from here. That’s why we’re giving 40 performance-hungry enthusiasts in the U.S. an opportunity to celebrate the next 40 years of high-performance computing by trading in their commemorative processor prize for our CPU that enables you to work, play and create with heavy metal.The first 40 U.S.-based winners of the Intel® 8th Gen Core i7-8086k Limited Edition Sweepstakes (which opened on June 7, 2018 and closed on June 8, 2018) (“Competitor Sweepstakes”) to complete certain steps will be offered the opportunity to exchange their new, in-box prize processor for a flagship 16-core AMD Ryzen™ Threadripper™ 1950X processor built for the gamers and creators who need a processor that can do it all. Those steps will be communicated on this webpage on June 25 at 1:00:00 PM EDT. Check back on June 25 at 1:00:00 PM EDT for complete details on how to participate.OPEN ONLY TO RESIDENTS OF 50 U.S./D.C. VOID OUTSIDE THE U.S. AND WHERE PROHIBITED.

AMD has a list of the complete terms and conditions on its site. But it is also noteworthy that "winners" of AMD's competing sweepstakes will have to pony up for a much more expensive X399 motherboard with the TR4 socket, which currently retail for more than $300, instead of Intel's less-expensive 300-series motherboards. Regardless, those who do swap their Intel Core silicon for an AMD Threadripper chip will gain 10 cores and quad-channel memory, not to mention quite a bit of resale value.

AMD suggests interested parties follow @AMDRyzen for further details. We'll be keeping a keen eye on Intel's response, but we don't expect it to up the ante with the $1,900 Core i9-7980XE.

EDIT: Intel responded via its official verified @IntelGaming twitter account, saying:

".@AMDRyzen, if you wanted an Intel Core i7-8086K processor too, you could have just asked us. :) Thanks for helping us celebrate the 8086!"

Building A Workstation With Ryzen Threadripper And Radeon Vega Frontier Edition

Igor Wallossek — Fri, 25 May 2018 01:00:00 +0000

Introduction & Components

Professionals who make money with their workstations need every bit of performance they can get. Overclocking, which often compromises long-term stability for a bit of extra speed, generally isn't an option. So, those folks need the fastest hardware available and it has to be more reliable than the average desktop. No wonder workstation components are so expensive! But when onerous price tags become a problem, it's sometimes acceptable to substitute prosumer hardware instead. Think high-end Core i9 processors and Nvidia Titan X or AMD Radeon Vega Frontier Edition graphics cards.

With last year's introduction of the AMD Ryzen Threadripper 1950X CPU, there's one more platform option for enthusiasts completing high-end tasks. It's relatively affordable, massively parallel thanks to a 16-core design, loaded with connectivity for add-in cards, and able to accommodate lots of memory. That makes Threadripper 1950X the perfect foundation for a value-oriented workstation build. And after we're done piecing this monster together, we'll use it for testing next-gen professional graphics cards.

Also included in the Tom's Hardware reference workstation is an ASRock X399-based motherboard, 32GB of DDR4 memory, AMD's aforementioned Radeon Vega Frontier Edition, a big Cooler Master Cosmos C700P case, and an Eizo EV3237-BK 4K monitor. Together, they give us plenty of performance, slots, and ports for a fast, stable PC adept at running professional software.

Motherboard And Memory

AMD's Ryzen Threadripper family utilizes a different processor socket than the company's mainstream Ryzen CPUs, which drop into the Socket AM4 interface. Rather, the more sophisticated Threadripper chips employ Socket TR4. And that means a new chipset to communicate with the processors, along with adding storage, USB, audio, and networking connectivity: X399.

ASRock's X399 Taichi offers us a good mix of moderate pricing (for a prosumer motherboard) and advanced functionality. It is not the most colorful, eye-catching, or well-equipped X399 board on the market. However, it's one of the most sensible for enthusiasts trying to keep the cost of their workstation manageable. Even under continuous load, the voltage regulation circuitry remains cool enough. And the absence of active cooling on the chipset's heat sink helps minimize noise.

As for memory, we used a total of 32GB DDR4 from two dual-channel Apacer Blade Fire kits. In this configuration, we're able to run the memory at 3200 MT/s after applying all of ASRock's BIOS updates. Trying to populate the remaining four memory slots necessitates a step down to 2866 MT/s, forcing you to choose between maximum capacity and optimal bandwidth. We ultimately went with the 32GB configuration because few of the applications we benchmark would have benefited from more RAM.

Readers in the U.S. may want to consider another brand instead, since Apacer isn't widely available domestically. We had good luck with G.Skill's Ripjaws V DDR4-3200 kit in our launch review, so a couple of 16GB kits would be a good alternative.

We could have gone more extreme with our storage, but again, most of our benchmarks stress the host processor and graphics card. For now, a pair of 1TB Crucial MX300 SSDs and a WD Red 4TB are enough.

Graphics

AMD's Radeon Vega Frontier Edition lives between two worlds, serving professionals and gamers at the same time. Whereas Nvidia's competing Titan Xp does not support the company's certified Quadro drivers, which would optimize it for applications like AutoCAD, 3ds Max, Maya, SolidWorks, etc., AMD enables the Frontier Edition card with its Radeon Pro Software drivers.

This is a critical differentiator for professionals. In some cases, workstation-class software won't even start without certified graphics drivers installed. By building this system with a Radeon Vega Frontier Edition, we're able to use a lot of applications at a much higher performance level than we'd get buying a Quadro for the same price.

Case And Power Supply

A powerful workstation calls for a case with an equally commanding presence. Since we have to move this machine from time to time, depending on whether we're testing performance, acoustics, or thermals, we chose Cooler Master's handy Cosmos C700P. On one hand, this case allows us to swap hardware for our tests quickly and easily thanks to its hinged side walls. On the other, we're positively surprised by its torsional rigidity. Good build quality and an attractive aesthetic add to the Cosmos' appeal.

The Cooler Master V850 does everything we need as well. It's based on Seasonic's mature KM3 platform, offers fully modular cabling, and sports stable voltage output.

Configured the way we have it now, our workstation requires a maximum of 600W during the most taxing stress tests we run. That leaves a bit of headroom before maxing out the V850's peak output. In a workstation environment, this is certainly desirable.

Monitor

Of no consequence to performance, but still critical to a great workstation experience, we went with an Eizo EV3237-BK display. The 32-inch IPS panel offers a native resolution of 3840x2160, along with the color space needed for CAD/CAM work. If your application is more color-sensitive, Eizo's ColorEdge line-up may be more appropriate.

Reference System
Processor	Ryzen Threadripper 1950X
Motherboard	X399 Taichi
Memory	32GB Apacer Fire Blade DDR4-3200 (4x 8GB)
Graphics	AMD Radeon Vega Frontier Edition 16GB
Hard Drive	Western Digital Red 4TB
SSDs	2x MX300 SSD 1050GB
Case	Cooler Master Cosmos C700P
Cooling	Floe Riing 360
Power Supply	Cooler Master V850
Monitor	Eizo EV3237-BK
OS	Windows 10 Pro 64-bit
Drivers	Radeon Pro Software

MORE: Best PC Builds

MORE: How To Build A PC

MORE: All PC Builds Content

Assembly & Testing

If you want to build a PC using Cooler Master's Cosmos C700P, you need a thick wallet and even thicker biceps. It's a beast of a chassis. Because of that, you'll probably want to lay down a soft blanket, then disassemble the side doors and cover during the build process.

If you're looking to replace Cooler Master's stock case fans with your own, be warned: the enclosure's mounting holes don't line up well with aftermarket 120mm fans. Even the screws bundled with Cooler Master's own MasterFan Pro won't work. We'll shelve this mod for now and come back to it later with suitable machine screws, nuts, and rubber grommets.

The closed-loop liquid cooler we used maxes out what the Cosmos C700P can accommodate. If you try to install a radiator longer than Thermaltake's 360mm Floe Riing, it won't fit. As you might imagine, then, this case may not be the best choice for a more elaborate custom water-cooling loop.

From there, all of the other components fit nicely inside of Cooler Master's case. Plastic shrouds covering the power supply and storage make the finished product look particularly clean. You just have to take the time to remove and reinstall them during the build process.

The Cosmos C700P is large enough to support multiple installation configurations: a regular layout, a chimney layout with the graphics card oriented vertically, and an inverted layout built upside down. Thermally, the regular layout is optimal due to the way case fans blow air through the case.

Benchmark Results

AMD's X399 platform and its Radeon Vega Frontier Edition both had rocky beginnings, requiring a series of software updates before they were truly considered mature. Today, the graphics card offers better performance, while Threadripper-based platforms support higher DDR4 transfer rates.

The ability to push more bandwidth through Ryzen Threadripper's quad-channel memory bus has a big impact on our benchmark results. Again, this is a workstation, so we're not interested in overclocking it. But a run through our 3ds Max 2015 workload shows how the CPU's performance is up substantially compared to last year's launch numbers.

The same goes for the Blender test, although the speed-up is a little less dramatic. Still, it's clear that AMD's firmware improvements benefit performance. Intel's competing Core i9s are simply blown away.

Early adopters updating to the latest BIOS versions and running their memory at higher data rates enjoy an almost 4% speed-up in Cinebench.

Isolating the graphics card, we compare AMD's Radeon Vega Frontier Edition with the latest Radeon Pro Software Adrenalin Edition driver to its performance at launch (and against Nvidia's Quadro P6000).

These benchmark results are slightly lower than what we measured in our AMD Radeon Vega Frontier Edition 16GB Review last year. That's because we were using an overclocked Intel Core i7-6900K-based platform back then. But AMD's Ryzen Threadripper 1950X still allows both high-end graphics cards to perform respectably.

Finally, we compare the relative performance of Radeon Vega Frontier Edition in its current and launch states to Nvidia's Quadro P6000, which is normalized at 100%.

Depending on the sub-composite score we look at, the Radeon Vega Frontier Edition's growth is notable. Though Nvidia's card is generally faster, the gap between them is shrinking. Besides, the Quadro costs several times more than AMD's solution.

Conclusion

Even a year ago, AMD wasn't able to compete in the workstation space with its CPUs, and the company's graphics cards were falling further behind Nvidia's professional boards based on the Pascal architecture. But with the introduction of Ryzen, particularly the 16-core Threadripper 1950X flagship, AMD has a host processor with enough performance to rival Intel's Core i9-7900X. The same goes for AMD's Vega GPU. It doesn't usurp Nvidia's competing Quadro P6000, based on the GP102 processor. However, it does come close, and at a much more palatable price.

Clearly, there are still scenarios where Intel's top-end Core i9s are quicker than the fastest Ryzen Threadripper CPU. Same goes for Nvidia's Quadro against the Radeon Vega Frontier Edition. But when it comes to performance for the price, AMD's platform cannot be beaten. And for that reason, we're eager to put this strong combination of parts to work doing what it does best: tearing through threaded tasks.

What's the overall damage for this monster of a workstation? Let's tally it all up:

Pricing
CPU	AMD Ryzen Threadripper 1950X	$880
Cooling	Thermaltake Floe Riing RGB 360	$175
Motherboard	ASRock X399 Taichi	$340
Memory	32GB G.Skill Ripjaws V DDR4-3200 (two 16GB kits)	$340
Graphics	AMD Radeon Vega Frontier Edition	$930
SSD	2x Crucial MX500 1TB SSD (MX300 no longer available)	$500
Hard Drive	WD Red 4TB	$125
Power Supply	Cooler Master V850	$190
Case	Cooler Master Cosmos C700P	$290
Total		$3770

MORE: Best PC Builds

MORE: How To Build A PC

MORE: All PC Builds Content

Gigabyte X399 Aorus Gaming 7 Motherboard Review: Threadripping In Style

Jacob Terkelsen — Thu, 03 May 2018 13:00:00 +0000

Features & Layout

Think of it as a jolt of Joe in motherboard form: Gigabyte’s X399 Aorus Gaming 7 delivers a bit of desktop-PC pick-me-up for days when your workflow is a snooze. Some might call this motherboard's profusion of LEDs downright gaudy, but our concerns are more practical: Can this Aorus board keep our reviewer grounded in performance, while elevating his prosumer status to 11 with loads of IO and pretty lights?

Spring is in the air, and the Gigabyte X399 Aorus Gaming 7 keeps our desk space bright. It's downright festive, with a shower of patterned lighting emitting from our Corsair 570X Threadripper test bench. Though AMD's Ryzen Threadripper chips target the prosumer market, the expectations of those buyers have come a long, long way from beige cases and Zip disks. Both the ASRock and MSI Threadripper boards we reviewed before this one provided LEDs enough for a window-chassis PC build, but Gigabyte kicks it up a notch with this Aorus-branded offering.

Everything about this product shouts "high-end." Though the retail box is smaller than those of the other two X399 review samples we mentioned and has a pleasant textured top and side panels, the bottom is festooned with marketing terms and feature descriptions. Opening the top greets us with a hard-plastic packaging support, as well as some motivational text for the gamer in all of us. Lifting the plastic tray unveils the bare board, with accessories below a cardboard tray.

Gigabyte does not go overboard with included goodies: a standard backplate, an SLI bridge, an Aorus badge, the usual G-Connector, and a couple of tie-down straps. (The company could have left out the graphics placard, since we're not VIPs, nor do we require "No Trespassing" signs in our work area.) The included Wi-Fi antenna uses long leads and a stand-alone antenna that you mount separately from the system chassis. Two thermistors are included, for additional in-chassis thermal monitoring, and two RGB-header extension cables are included to help wire out custom lighting for your build.

As for lighting, simply placing a few LEDs by the X399 heatsink won’t cut it for an Aorus product, it seems. When this board is powered on, customizable lighting engulfs the X399 heatsink, the I/O shroud, and the PCB's audio-circuitry zone while surrounding the PCIe x16 and memory DIMM slots. The lighting scheme might look a bit obnoxious when staring at product shots or box art, but when combined with a fully populated system, the LEDs do accent the components on the board rather than being the star of the show. If the included lighting does not scratch your RGB itch hard enough, three RGB headers on the board may help: two at the bottom left, and one at the top right. (All this lighting gets controlled via Gigabyte's RGB Fusion software, which you can access through the UEFI BIOS or a Windows front-end.)

The back panel for the Gigabyte X399 Aorus Gaming 7 is a good mix of old and new. A single PS/2 port is available for a (very) legacy keyboard or mouse, along with two DAC-UP-enabled USB 3.1 Gen1 ports. Six USB 3.1 Gen1 ports are below that, along with a Gigabit RJ-45 jack. (More on that in a bit.) For the Wi-Fi card, two MMCX connectors are mounted for the included antenna accessory. Back-panel USB 3.1 Gen2 is present in the form of one Type-A and one Type-C connector. For audio, five gold-plated analog jacks and an optical output enable the ALC1220 codec and Gigabyte’s AMP-UP technologies, featuring Sound Blaster 720˚.

So, what’s the deal with this Gigabit Ethernet controller? Gigabyte partnered with Rivet Networks to embed the latter's Killer E2500 NIC in this board for an improved network experience compared to the Intel standard offerings. We took some time in 2017 to learn more at Rivet’s Austin campus, and we received a detailed briefing as to what this controller is capable of. In the lab, we were able to observe real latency reductions while playing CS:GO, BitTorrenting Ubuntu images, and watching 4K YouTube on a separate monitor. Though we're not networking engineers, we can more or less reproduce Rivet's results when plugged in to a standard Gigabit Ethernet-capable router and broadband service. As for the Gigabyte X399 Aorus Gaming 7, the inclusion of this NIC is a good effort at differentiating from the competition and well worth some of the extra cost.

One nice departure from other boards we’ve reviewed is Gigabyte’s choice of spreading the remaining headers to the outer edges of the board, rather than sneaking them between major components (except for one of the fan headers). The bottom edge of the board hosts headers for front-panel audio, TPM, USB 2.0 (two of them), USB 3.0, and the front-panel connections. Reset, Clear CMOS, and Power buttons are included on the bottom edge directly below the last PCIe slot, so using them in conjunction with dual-slot cards at the bottom will pose a challenge. The numeric debug readout is to the left of the front-panel connector, and two four-pin fan headers are available here.

The right side of the board contains eight angled SATA3 ports and access to a front USB 3.1 Gen2 header. Above that is the usual 24-pin ATX connector, as well as three additional four-pin fan headers. Going topside, eight-pin and four-pin 12V ECS plugs are squeezed into the top left side of the board. This segment of the board is very cramped, with the routing of the Vreg heat pipe along with sizable heatsinks, DDR4 DIMMs, and the last remaining fan header, for a total of eight fan headers.

The inner portions of the board depart from the symmetry of the MSI X399 Carbon Pro Gaming and reflect a layout closer to the ASRock Taichi. Five PCIe x16 slots are available and routed out from top to bottom as x16, x8, x4, x16, and x8. A nice touch: The lane-bandwidth designations are printed on the board, so we do not have to trudge through the manual to investigate. Two full-length (110mm) M.2 slots are placed below and above the top and bottom PCIe slots, respectively, and a shorter (80mm) M.2 slot resides directly below the X399 heatsink. These M.2 solutions also employ Gigabyte’s Thermal Guard SSD heatsinks, with thermal padding to help spread heat across the SSDs' surface area, rather than leaving those slim storage sticks to fend for themselves.

The layout is nearly ideal, but we do have some quibbles with the design. As we were photographing this specimen, we had to remove the bottom M.2 heatsink to show the hidden CMOS battery. Unfortunately, the retention screw was torqued down so firmly that it was spinning the metal spacer. With a set of pliers, we could remove the screw, but operating within that tight space with a tool might make you damage the red audio capacitors. Be careful!

Also, as we mentioned in our first X399 board review, it is never wise to apply a person’s full body weight to the retention mechanism of any socket. However, we can confirm that the Foxconn retention mechanism used here does not fasten the CPU as straightforwardly as our other two review specimens. By changing the retention-screw order, we were able to get the Foxconn socket to fasten without additional force.

Also, did we mention that the EPS 12V plugs are very cramped? Luckily, the latches are closer to top of the board, so they are accessible and our thumbs don’t bump into unnecessary shrouding when trying to squeeze them.

MORE: Best Motherboards

MORE: All Motherboard Content

UEFI

The Gigabyte UEFI is a spitting image of what we saw in our X370 coverage and only differs with respect to Threadripper. The main screen greets us with a lightly stylized frame with our standard options residing in the blackness of the page.

Pressing the right Alt key will open the CPU, memory, and voltage health stats for the system. Pressing the left Alt key will open the help menu with a brief description of the field and some tabs for “Easy” mode, language, Q-Flash and Smart Fan.

Along the top, Motherboard Intelligent Tweaker (M.I.T.) contains most of the standard overclocking submenus. By default, frequency options place the user into Auto mode, and the user can either key in the new multiplier or press the “+” key. The desired frequency will populate the left column for this row, and the current setting will show up to the right. Similarly, XMP can be enabled for however many SPD modes the DIMMs support, and you can increase the memory multiplier with the same presses of the keyboard. Advanced CPU Core Settings gives access to features such as enabling Core Performance Boost, Cool & Quiet, and SMT mode. Hopping back to the main M.I.T. page, you can more heavily tune and tweak memory configurations for tighter timings. PC Health settings is redundant, but it gives you another panel for viewing various motherboard-reported voltages.

Switching gears to Smart Fan, we are greeted by a well-designed UI, compared to some of our other review samples' efforts. The Monitor pull-down refreshes the contents of the screen with information for that specific fan header and its current configuration profile. It's possible to customize each fan’s operation and tracking sensor.

RGB Fusion, in contrast, is a very basic interface with limited options for color selection and lighting patterns. We preferred the static or pulse mode with this board, as the flash and double-flash mode make the board look like a failing fluorescent bulb. However, when combined with other RGB and LED strips, the Windows-based application can orchestrate some impressive patterns and synchronize the entire build.

With M.I.T. out of the way, the peripherals menu is useful for tweaking the lower-level AMD CBS, Zen features, and even the DAC-UP feature (the yellow USB ports on the back). Diving into the lower-level Zen options brings us back to the “No help string” descriptions and other OC modes defined by the Zen architecture.

Overall, the UEFI could use a little more character, but it is completely functional, and the feature descriptions are clear. IOMMU appears to be disabled by default, which was a thorn in our side in the course of previous X399-board reviews. (If there are any additional features you want us to go into further detail, please let us know down in the comments.)

MORE: Best Motherboards

MORE: All Motherboard Content

How We Test

The Corsair Crystal Series 570X still houses our Threadripper test bed, and it pairs well with the Gigabyte X399 Aorus Gaming 7, as far as aesthetics are concerned. This previously reviewed chassis screams "sex appeal!" with tempered glass on four sides, as well as three prominent 120mm LED fans in front (with room for 360mm or 240mm radiators for liquid cooling). Since it is a review rig, we’ve been moving components in and out quite a bit, so the rubber pads for the glass standoffs have started to erode. Also, dust isn’t too much of a concern for us, but the case has plenty of gaps throughout at the glass joints, so keep a can of compressed air handy.

An Elgato HD60 Pro still lives comfortably in one of our PCIe x16 slots, waiting for the chance to stream some good PUBG, CS:GO, or general benchmark tests. As for the remaining components, there are no changes to our rig. Antec’s 1200W High Current power supply provides the juice to our Gigabyte GTX 1080 G1 Gaming GPU and Ryzen Threadripper 1950X CPU. G.Skill provides the 32GB of DDR4-3200MHz RAM, and Toshiba still supports all of our motherboard reviewers with its 256GB RD400 M.2 NVMe drive. Fractal Design’s Celsius S36 360mm water cooler is placed on the intake side of our chassis to provide cooling for our 190W processor.

Test Hardware

Benchmark Settings
Synthetic Benchmarks & Settings
PCMark 8	Version 2.7.613 Home, Creative, Work, Storage, Applications (Adobe & Microsoft)
SiSoftware Sandra	Version 2016.03.22.21 CPU Arithmetic, Multimedia, Cryptography Memory Bandwidth
Iometer	Version 1.1.0 4K Random Read, 4K Random Write 128K Sequential Read, 128K Sequential Write
Cinebench R15	Build RC83328DEMO OpenGL Benchmark
CompuBench	Version 1.5.8 Face Detection, Optical Flow, Ocean Surface, Ray Tracing
3D Tests & Settings
3DMark 13	Version 4.47.597.0 Test Set 1: Skydiver, 1920x1080, Default Preset Test Set 2: Firestrike, 1920x1080, Default Preset Test Set 3: Firestrike Extreme, 2560x1440 Default Preset
Application Tests & Settings
HandBrake CLI	Version: 0.9.9 Sintel Open Movie Project 4.19GB 4K mkv to x265 mp4
LAME MP3	Version 3.98.3 Mixed 271MB WAV to mp3 Command: -b 160 --nores (160Kb/s)
Adobe After Effects CC	Release 2015.3.0 Version 13.8.0.144 PCMark-driven routine
Adobe Photoshop CC	Release 2015.5.0 20160603.r.88 x64 PCMark-driven routine (light and heavy)
Adobe InDesign CC	Release 2015.4 Build 11.4.0.90 x64 PCMark-driven routine
Adobe Illustrator	Release 2015.3.0 Version 20.0.0 (64-bit) PCMark-driven routine
Blender	Version 2.68a BMW 27 CPU Render Benchmark BMW 27 GPU Render Benchmark
7-Zip	Version 16.02 THG-Workload (7.6GB) to .7z, command line switches "a -t7z -r -m0=LZMA2 -mx=9"
Game Tests & Settings
Ashes of the Singularity: Escalation	Version 2.50.28527 High Preset - 1920x1080, Mid Shadow Quality, 1x MSAA Crazy Preset - 1920x1080, High Shadow Quality, 2x MSAA High Preset - 3840x2160, Mid Shadow Quality, 1x MSAA Crazy Preset - 3840x2160, High Shadow Quality, 2x MSAA
F1 2015	2015 Season, Abu Dhabi Track, Rain 1920x1080 - Ultra High Preset, 16x AF 3840x2160 - Ultra High Preset, 16x AF
Metro Last Light Redux	Version 3.00 x64 High Quality, 1920x1080, High Tesselation Very High Quality, 1920x1080, Very High Tesselation High Quality, 3840x2160, High Tesselation Very High Quality, 3840x2160, Very High Tesselation
The Talos Principle	Version 301136 1920x1080 - High Preset, High Quality, High Tesselation, 4x AF 1920x1080 - Ultra Preset, Very High Quality, Very High Tesselation, 16x AF 3840x2160 - High Preset, High Quality, High Tesselation, 4x AF 3840x2160 - Ultra Preset, Very High Quality, Very High Tesselation, 16x AF

MORE: Best Motherboards

MORE: All Motherboard Content

Benchmark Results & Final Analysis

We now have three proper X399 boards to compare against one another, so we will remove the Gaming Mode metrics from our charts. From the brief validation we did with this product, Gaming Mode still provides about 30% less performance for Threadripper in this configuration.

Comparison Products

ASRock X399 Taichi

Gigabyte X399 Aorus Gaming 7

MSI X399 Gaming Pro Carbon AC

Benchmarks: Synthetics

PCMark shows some mixed results for the Gigabyte X399 Aorus Gaming 7. For Home, Work, and Storage, the Aorus is an average performer, but both Adobe and Office show sizable differences compared to the X399 Taichi, which is our PCMark winner. Sandra Arithmetic, Multimedia, and Memory Bandwidth place the Aorus slightly above the sample average across metrics, and the Cryptography benchmark favors the Aorus by the smallest of margins.

Sequential-read and -write results for the Toshiba RD400 in our X399 test benches are consistent, with just a 1.5 percentage point spread across motherboards. On the other hand, random reads are spread by about 6%, and writes see a sizable 10% delta, with the X399 Aorus winning on reads and losing on writes. Regardless, we saw a virtual tie across all three boards when combining our analysis with the PCMark storage metric.

The X399 Aorus’ single-threaded performance in Cinebench is encouraging, with a 2.8% win against the MSI X399. The multi-threaded score continues the win streak, though this time by only 0.5% against the Taichi. The more GPU-centric workloads in Compubench also favor the Gigabyte X399 Aorus Gaming 7, with a modest 1% win against the Taichi with all workloads averaged together. In fact, the Aorus wins in all Compubench metrics except Optical Flow, Bitcoin, and FFTs.

The X399 setup that we use for our motherboard benchmarking is complete overkill for 3DMark Skydiver. Regardless, the ASRock X399 is lagging by quite a bit, while the Gigabyte and MSI boards are performing more in line with expectations. The situation changes with Firestrike, where the ASRock takes the lead and the graphics delta between the Gigabyte and ASRock is nearly 2%. Firestrike Extreme shows the smallest delta in the metric, with just 0.25% separating the top and bottom performers. Physics performance, on the other hand, is more dramatic, with the Gigabyte X399 Aorus Gaming 7 taking the "win" by 0.35%.

Benchmarks: Applications

Application metrics can be misleading when comparing percentages to time results, if you're not careful to look at the measures in play. Where one product succeeds by a large percentage, say, another might lose by only a few seconds.

Our first example, the Handbrake metric, favors the MSI X399 Carbon by 1.8%, but that translates to only 27 seconds in practical fact. In the grand scheme of the workload, that’s less than 2% of the actual benchmark itself. The Aorus board lags behind by only a half-second on average in Adobe, in Blender GPU by 6 seconds, and in 7-Zip by a single second. (We tried to adjust our scales in the charts here to maintain a sense of balance: both representing the differences between products and keeping true to the larger context of the workload.)

Benchmarks: Games

Ashes of the Singularity: Escalation is probably our best representation of how X399 behaves inside of our motherboard bench suite. The Gigabyte X399 Aorus Gaming 7 leads the pack by 0.6% at 1080p and high details, and it squeaks ahead at the Crazy detail setting by less than a tenth of a percent. Running the tests at 4K with this game does reduce frame rates a bit, and the Aorus board was mid-pack in these results by a smidge.

The Talos Principle appears to have stabilized a bit (when compared to what we saw in our AMD AM4 reviews), and variances between products are roughly 5 to 10 frames per second. At 1080p and Ultra settings, the Aorus comes ahead by 1.1% compared to the Taichi, thanks to its better minimum and average frame rates. At 4K and High settings, the Aorus leads the pack by 3 frames per second compared to the Taichi, and Ultra settings bring each of the products within 2 frames per second of each other across all metrics.

In Metro Last Light, we see some big frame-time differences across test runs for each of these products, so seeing a minimum-to-maximum difference of up to 167 frames is quite concerning. Fortunately for us, we only factor in average frame rates as generated by the benchmark utility in our calculations, so at 1080p and high settings, the Aorus X399 performs admirably compared to the Taichi. Increasing the quality settings to Very High helps smooth out the frame graphs and as such, average rates are more representative.

F1 2015 is our final games workload, and we have to consider that this application cannot run with all 32 threads activated. Windows Resource Monitor reports that the game is activating only up to 22 threads. CPU threads aside, Ultra settings are smooth and enjoyable, and the Gigabyte X399 Aorus Gaming 7 comes ahead in terms of average and maximum frame rates, while minimum frames are a bit behind the pack. Testing at 4K doesn’t improve things for the Gigabyte board, showing a 3-frame-per-second loss to the MSI X399 Gaming Pro Carbon.

Power, Thermals, Performance & Efficiency

We keep a Kill-A-Watt connected to the wall and still report overall system power draw factoring in all components. At idle (with 4K monitor enabled), the Gigabyte X399 Aorus Gaming 7 shows an average result. CPU load, however, shows this system drawing 11 extra watts from the wall. GPU tests again put it in the average result category, but overall system stress balloons by nearly 51W. Double-checking our logged data and re-running the tests do confirm that the results are repeatable; this was recorded under the most extreme FurMark test (in 4K Windowed mode, and with 4x MSAA at ~97% utilization) and our highest-load instance of Prime95 (with extended run times showing Prime95 cycling between states). Ouch.

As for thermals, the Fractal Design Celsius S36 in intake configuration performs just as well with the Aorus X399 compared to the ASRock X399 board we tested. However, the Vreg heatsink design on this board seems less effective, which could just be an artifact of the processor pulling more amperage during load condition. Either way, all three boards and configurations are prime gear for overclocking.

Performance-wise, all three products are within 1.5% of each other across all the metrics. The Gigabyte X399 Aorus Gaming 7 performs well in the games and synthetic portions, but the apps suite favors the test times of the Taichi by roughly 1.4%.

Even if we shave 40W off the maximum load condition for the Gigabyte X399 Aorus, this board simply can’t beat the efficiency of the ASRock X399 Taichi. The performance metrics for all three of these boards are very close, so any extra watt lost is a fraction of a percent in the wrong direction. Still, the Aorus X399 performs well and as we start to overclock, maybe we’ll see that extra power draw pay off.

Overclocking: 4GHZ, Finally!

Single DIMMs plugged in per channel provides for the easiest overclocking when using our test G.Skill Trident Z DDR4-3200MHz kits. As with the other three boards, DDR4-3333MHz is easily achievable and highly stable throughout the test suite. (Maybe it’s time to ask G.Skill or Corsair to send some DDR4-4000 kits for evaluation?) In terms of timings, the advertised CL14 is working with the default XMP profile at the increased frequency. It might be worth undervolting these DIMMs to shave off a few extra watts, if you're so inclined.

Core overclocking proves straightforward on this board for the Threadripper 1950X sample we have. Simply dialing up the UEFI to a 38.00 multiplier provides for an easy 400MHz boost compared to the stock 3.4GHz speed. Getting to 3.9GHz does require increasing voltage to 1.25V in the UEFI. 3975MHz is achievable if tuning up the load line to High. Bumping up the voltage up to 1.2875V gave us our target of 4GHz, and that while running Prime95 for an extended time. However, this processor is cooking at ~85° C Tdie and temperatures of about ~105° C for Tctl as reported by HWiNFO. This overclock was stable for running in extended AIDA64 tests at ~63° C, and it gave us no problems when streaming and encoding 4K input to 1080p.

Value, Conclusion & the Road Ahead

Strictly looking at MSRP, the Gigabyte X399 Aorus Gaming 7 falls short when compared to the more stripped-down ASRock X399 Taichi. Let’s run through the extra features, though, to help justify this cost.

The Killer E2500 is a bonus feature that alone can help any bandwidth-restricted user gain additional perceived performance increases, even if it doesn’t net any additional frames. The sheer number of fan and RGB headers, combined with their placement, is also ideal for visually-minded setups that want to run complex radiator arrangements without a fan controller. Though the UEFI isn’t as visually impressive as the ASRock offering's, the RGB Fusion, Smart Fan, and other UEFI features are, in our opinion, more user-friendly on the Gigabyte/Aorus product.

Tie that in with the cost of a Threadripper 1950X chip, the GeForce GTX 1080, and cooling hardware, and you start to see just how minor that $40 delta can be in a comparative sense. Shop around, look for sales and rebates, and that price keeps going down. The feature mix--and, if you're so inclined to use it, the profusion of lighting--are how the Gigabyte X399 Aorus Gaming 7 earns its top marks for an X399 board.

MORE: Best Motherboards

MORE: All Motherboard Content

How to Choose the Right CPU for 1080p Gaming: 14 Processors Compared

palcorn@outlook.com (Paul Alcorn) — Fri, 23 Mar 2018 18:30:00 +0000

Introduction

Last year, we saw innovation from AMD and Intel that we hadn't experienced in a long, long time. The Zen architecture made AMD competitive in segments of the CPU market it previously couldn't touch, and Intel moved as quickly as possible to defend its incumbent position. We thoroughly enjoyed the back-and-forth as both companies jockeyed for enthusiasts' adoration.

But even as new platforms were springing up with more PCIe connectivity than ever before, graphics-card availability dried up. Cryptocurrency miners bought up everything they could find to capitalize on rising valuations. Even today, you can't find modern models anywhere near their suggested retail pricing. We've resorted to buying pre-built systems and scouring the forums for previous-generation cards, trying to score a bargain.

The best spread of CPU technology in ages, paired with sky-high GPU prices, is a recipe for confusion for PC builders. For the same amount of money, enthusiasts can afford less graphics performance than they could not long ago. That makes it easy to overspend on host processing, since balance is thrown out of whack. But you can also get a lot more CPU for your dollar than this time a year ago. How do you make sure you're getting the most for your budget?

Well, we set out on a mission with 14 CPUs and three different GPUs to find the best combination in nine popular games.

Moving The Goalposts

For the last 11 years, Core i7 and Core i5 CPUs featured four cores. Coffee Lake changed this. Now, Core i3s sport four cores, Core i5s include six cores, and Core i7s boast six Hyper-Threaded cores. Intel also gave its low- and high-end models a makeover: Skylake-X stretches up to 18 cores/32 threads for high-end desktops, while Pentium processors have now gained Hyper-Threading technology.

Of course, AMD introduced its line-up of Ryzen 7, 5, and 3 models with copious core counts. Moreover, the Ryzen Threadripper series landed with up to 16 cores/32 threads and such friendly prices that Intel was forced to make its Skylake-X chips more affordable.

The Contenders

As you might imagine, the old rules of picking a CPU family to go with certain graphics cards changed as a result of these new processors. Thus, we decided to investigate using the best performers from each CPU class.

Representing AMD, we have the Ryzen 7, 5, and 3 models. We didn't bother testing last-generation Bulldozer-based CPUs, but we did throw in the Ryzen Threadripper 1950X for good measure.

For Intel, we have K-series Core i7, i5, and i3 CPUs from the Coffee Lake and Kaby Lake generations. We also added the Core i9-7900X and Core i9-7980XE to cover high-end desktops. Out of curiosity (or because we're gluttons for benchmarking punishment?), we couldn't help but include two of the latest Pentium processors, too.

That gives us 14 processors spread across five test platforms. We paired these with the GeForce GTX 1080, GTX 1070, and GTX 1060 (6GB) graphics cards. Although the GTX 1080 is considered an extravagance these days, we have to imagine it'll come down in price someday.

Finally, we selected nine games for testing. Some of the titles are new, while others are older. We did weigh the suite, though, more toward modern games. Some of them are CPU-dependent, others are decidedly graphics-bound, and a few are actually pretty well split down the middle. This allows us to explore bottlenecks from different angles.

Today's tests are all run at 1920x1080. (We have more data coming at 2560x1440 and 3840x2160, too, so expect a follow-up story or stories to present our findings there.) To best represent the experience we'd want to have, all benchmarks were run with the highest graphics settings possible.

Test Systems

To avoid variance from GPU Boost as our GeForce GTX graphics cards heat up, we use multiple runs from each benchmark in quick succession. We select the median value from the last recordings.

Test System & Configuration

Hardware

Intel LGA 1151 (Z370) @ 2666/2400AMD Socket AM4 @ 2667 Intel LGA 1151 (Z270) @ 2400AMD Socket SP3 (TR4) (In Game Mode) @ 2667 Intel LGA 2066 @ 2666 Common To AllGigabyte GeForce GTX 1060 G1 Gaming 6G 1TB 1500W (Creators Update Version 1703)

MORE: Best CPUs

MORE: All CPUs Content

Ashes Of The Singularity: Escalation

Nvidia GeForce GTX 1080

Ashes of the Singularity was one of the first titles to receive a Ryzen-optimized patch. And it appears that host processing is always going to be your bottleneck if you're using a GeForce GTX 1080 at 1920x1080. As we can see, the 18-core/36-thread Core i9-7980XE takes a healthy lead over the rest of the field, while the 10C/20T Core i9-7900X captures second place.

This benchmark scales well with increased core counts, highlighting some of the generational transitions in Intel's line-up. The 6-core/12-thread Core i7-8700K illustrates a big jump over Intel's previous-gen 4-core/8-thread Core i7-7700K, which basically ties the Core i5-8600K.

The Core i3-7350K (and every CPU below it in our chart) struggles mightily during the benchmark.

Nvidia GeForce GTX 1070

Stepping down a notch to the GeForce GTX 1070 finds Intel's Core i9-7980XE and Core i9-7900X providing nearly identical performance, while the rest of the field scales down as we'd expect.

Really, the Core i7-7700K and most processors under it offer nearly the same performance as they did paired up to a GeForce GTX 1080. We do measure a 2.3 FPS delta between the i5-8600K with both graphics cards, but that appears to be an exception. Such small variances hardly make it worth stepping up to the GeForce GTX 1080 in this title at this resolution.

Nvidia GeForce GTX 1060

The limitations of a GeForce GTX 1060 become clear, even in this CPU-bound test, though Intel's Core i9-7980XE still maintains a healthy lead over the rest of the field.

At this point, a graphics bottleneck is emerging, pushing the GTX 1060 to its limit. As a result, the Core i9-7900X and Core i7-8700K fall a few positions. It's possible that a lighter workload isn't pushing the execution cores hard enough for them to maintain their highest Turbo Boost bins. If true, the uncore frequency drops as well, slowing the cache and ring bus frequencies, reducing performance throughout. We'll see this phenomenon in other titles, too.

Compared to our GeForce GTX 1080 results, we see similar performance from every CPU beneath the Core i5-8600K, showing that lower-end processors are still the bottleneck.

MORE: Best CPUs

MORE: All CPUs Content

Civilization VI Graphics

Nvidia GeForce GTX 1080

Intel's Core i9-7900X leads this benchmark with an impressive 106.6 FPS. The remainder of the results are uneventful, generally following trends we've observed through past CPU reviews. The Core i7-7700K averages 102.8 FPS, and the nearest mainstream Ryzen competitor achieves 80.1 FPS.

We tested the Threadripper 1950X in Game Mode simply because there wasn't enough room in our charts for the many combinations that AMD enables. Head over to our Threadripper 1950X Game Mode, Benchmarked article if you're interested in the impact of the various modes on Threadripper's performance. Just be aware that the processor behaves differently in many games, depending on the settings you choose.

Nvidia GeForce GTX 1070

Stepping down to the GeForce GTX 1070 exposes a graphics bottleneck with some of our faster CPUs. Intel's Core i9 models push the limits at around 94 FPS, landing within the range we'd expect for a graphics-bound situation. The rest of the field falls into line based on host processing power, though some CPUs fare worse than others.

Nvidia GeForce GTX 1060

A more definite GPU limitation irons out some of the variance between CPUs. Worthwhile scaling is no longer observable. Case in point: There is only a 6.4% delta between the Ryzen 5 1600X and Threadripper 1950X, showing how a midrange graphics card can neutralize the benefit of a high-end processor if you aren't careful to maintain balance.

Fortunately, this also means you can snag an affordable Ryzen 3 or Core i3 and enjoy almost the same performance as you would with a much more expensive model.

MORE: Best CPUs

MORE: All CPUs Content

Dawn Of War III

Nvidia GeForce GTX 1080

Dawn of War III scales well with increased execution resources. As we can see from the fastest CPUs, though, it also favors high clock rates.

Intel's Core i7-8700K reaches up to 113.6 FPS. However, its 1.5 FPS advantage over the Core i7-7700K isn't very impressive considering the Core i7-8700K boasts 50% more Hyper-Threaded cores.

Meanwhile, the $1000 Core i9-7900X doesn't offer as much performance as a $175 Core i3-8350K. This is why we don't recommend HEDT processors for purpose-built gaming rigs.

Nvidia GeForce GTX 1070

The same general scaling trends persist with the GeForce GTX 1070, albeit at a lower performance level. Intel's Core i3-8350K continues behaving like older Core i5 CPUs, while the Core i5-8600K really hits a price/performance sweet spot with this pairing.

Nvidia GeForce GTX 1060

Again, some of the highest-end processors drop a few places because they aren't utilized fully. There's only a 3.3 FPS (4%) delta between the Core i9-7980XE and the first-place finisher...but that chart-topper is a Core i5!

Again: Striking the right balance is important. Not only is it a waste to buy too much CPU if you don't need the processing power for other workloads, but you can even handicap it with a graphics card incapable of keeping it busy.

MORE: Best CPUs

MORE: All CPUs Content

Far Cry Primal

Nvidia GeForce GTX 1080

Far Cry Primal's Dunia Engine 2 responds well to high clock rates and IPC throughput. However, it also benefits when we turn off simultaneous multi-threading. As a result, you'll notice that the Core i5s are very competitive.

An increasing number of games are being tuned for quad-core CPUs, and this one is no exception. We think it'll take some time before chips with more cores start distinguishing themselves, though Coffee Lake fares well enough.

Although AMD doesn't do as well in our average measurements, its Ryzen 5 1600X and Ryzen 7 1800X processors demonstrate the least amount of frame-time variance. Rest assured that they offer a smooth gaming experience.

Nvidia GeForce GTX 1070

We observe much less variance between CPUs as we step down to a less powerful GeForce GTX 1070. The 6-core/12-thread Ryzen 5 1600X continues outmaneuvering the 8-core/16-thread Ryzen 7 1800X, while the Threadripper 1950X inexplicably jumps higher in the rankings.

Nvidia GeForce GTX 1060

We enter graphics-bound territory with the GeForce GTX 1060, which mixes up the finishing order. Notably, even though they don't lead when it comes to average frame rate, Intel's Core i7-8700K and i5-8600K provide the smoothest performance with the least frame-time variance. The Pentium models also respond well to this pairing.

AMD's Threadripper 1950X is the only processor that lands below our expectations. Then again, we've proven that Game Mode isn't the best option for this title. A bit of tuning might rectify the issue.

MORE: Best CPUs

MORE: All CPUs Content

Grand Theft Auto V

Nvidia GeForce GTX 1080

Grand Theft Auto V favors Intel architectures and, more generally, multi-core designs with high clock rates. It may seem odd that the Core i5-8600K beats Intel's Core i7-8700K, but we've observed this tendency on other occasions.

Did you notice that Core i3-8350K out there in front of the quad-core Core i5-7600K? Clearly, Coffee Lake propels the Core i3 family to new performance heights.

Nvidia GeForce GTX 1070

The finishing order remains intact as we test with Nvidia's GeForce GTX 1070, though average frame rates fall a bit. Still, many of the slower models land within 1 FPS of what we saw from the GTX 1080, suggesting that we're not seeing a graphics-bound workload. Host processing appears to be the limiting factor.

Nvidia GeForce GTX 1060

Plain and simple: Nvidia's GeForce GTX 1060 restricts performance at this point.

MORE: Best CPUs

MORE: All CPUs Content

Hitman

Nvidia GeForce GTX 1080

We began our testing with the Pentium CPUs and worked our way up, benchmarking all nine games for each processor before moving on to the next. So, you can imagine how thrilled we were when we discovered that IO Interactive implemented a 90 FPS frame cap at some point through a patch. Soldiering on anyway, we observed little scaling above the Core i9-7900X.

Intel's Pentium processors average a solid 60+ FPS, though we could see quite a bit of hitching during the test sequence. The Ryzen 3 1300X and the Core i3-7350K fare better, but you want a Ryzen 5 or a Kaby Lake-based Core i5, at the least, for smooth performance in this title.

Nvidia GeForce GTX 1070

The frame cap continues to be a killjoy, but at least a GeForce GTX 1070 and a reasonably-priced Core i5-8600K push Hitman's limits at 1920x1080. There is a bit of unevenness to report from the Ryzen 3 1300X and Pentium models, which manifests in our frame-time variance chart.

Nvidia GeForce GTX 1060

We finally duck under the frame cap, but only because Nvidia's GeForce GTX 1060 isn't fast enough.

There's only a 3% delta between the quad-core Core i3-8350K and first-place finisher. This is definitely a graphics-bound workload that only breaks apart a bit when we drop to the lowest-end CPUs.

MORE: Best CPUs

MORE: All CPUs Content

Shadow Of War

Nvidia GeForce GTX 1080

Middle-earth: Shadow of War doesn't scale as dramatically as some of our other benchmarks, and it certainly isn't as sensitive to IPC throughput and frequency as Shadow of Mordor. CPU reviews tend to focus on games that scale well with certain host processing specifications, such as core count, clock rate, cache size, or memory bandwidth. But some games just can't get enough graphics performance.

This title exhibits a 12.6 FPS (~16%) gap between the fastest and slowest processors in our pool, but every system surpasses the 70 FPS mark easily. There is a difference in smoothness between the Pentium and Ryzen 3 1300X processors, but it isn't extreme.

Nvidia GeForce GTX 1070

Naturally, a GeForce GTX 1070 is even more tied up, tightening the delta from top to bottom to just 3.3 FPS.

The Pentium processors still struggle a bit with hitching.

Nvidia GeForce GTX 1060

A GeForce GTX 1060 puts all of these CPUs on an even footing. Subjectively, there is no difference between the processors, even though our data suggests that the Core i7-8700K lags behind.

MORE: Best CPUs

MORE: All CPUs Content

Project CARS 2

Nvidia GeForce GTX 1080

Project CARS 2 is purportedly optimized for threading, and it seems to scale well with increased host-processing resources. After all, we see a 41 FPS delta between the fastest and slowest processors in our test pool. The Core i5-8600K leads, and it is followed closely by Intel's Core i7-7700K.

Nvidia GeForce GTX 1070

The Core i3-8350K shows that it's a surprisingly capable quad-core CPU, besting AMD's entire line-up. With that said, the Ryzen 5 1600X specifically offers most of what higher-end models can do, but at a much lower price.

Nvidia GeForce GTX 1060

The GeForce GTX 1060 limits performance at 1920x1080. AMD's Ryzen 3 1300X is the only processor to exhibit a marked amount of hitching during our recording.

MORE: Best CPUs

MORE: All CPUs Content

PlayerUnknown’s Battlegrounds (PUBG)

Nvidia GeForce GTX 1080

PlayerUnknown's Battlegrounds can be a frustrating title to benchmark. Its constant cadence of updates seems to introduce more problems than are solved, and we've seen visible hitching from even the fastest CPUs and GPUs.

Aside from the Pentiums, scaling in PUBG is fairly meager. AMD's Ryzen 3 1300X suffers the most jarring stuttering during playback of our recorded game.

Nvidia GeForce GTX 1070

Stepping down to a GeForce GTX 1070 confirms that we're decidedly graphics-bound. Not that this game is particularly detailed; it may just be poorly optimized. Either way, there's no real meaningful conclusion we can draw from such an outcome.

Nvidia GeForce GTX 1060

Our final test leaves little to explain: There are no meaningful differences between CPUs when you're constrained by a GeForce GTX 1060.

MORE: Best CPUs

MORE: All CPUs Content

The Best Combinations of CPU and GPU at 1080p

Boiling our testing down into one general recommendation is almost impossible, especially given the pricing of graphics cards right now. But we can identify some of the more obvious trends.

Fast-paced first-person shooters tend to prefer high clock rates and fast graphics cards. Nimble CPUs like the Intel Core i5 and i7, and AMD's Ryzen 5 and 7, fare best in our performance measurements. Some titles are easily graphics-bound, though that could be due to poor optimizations or minimal effort to parallelize the game engine. PlayerUnknown’s Battlegrounds is the perfect example. It doesn't respond at all to faster processors. It's far better to buy a higher-end graphics card if that game is your main squeeze.

Real-time strategy titles seem to respond more favorably to CPUs with lots of cores; so do many turn-based games. Of course, that genre isn't as sensitive to graphics performance because it doesn't require quick response times. But a fluid experience in RTS games is a must. Often, host processing is the bottleneck, so dialing back your GPU investment and buying a better CPU yields the best performance possible.

Charting Out The Trends: The Best Pairings

We typically plot our test results on a scatter chart with price and performance on two axes, but we aren't even going to take a stab at graphics pricing during these tumultuous times. Instead, we're providing a geometric mean of the 99th percentile frame times, representing smoothness, converted into an FPS measurement for each class of graphics card. Each game factors into this calculation except for the artificially hamstrung Hitman.

Balance is the goal here. Intel's Core i7-8700K is the first chip we think of when it comes to top-of-the-line for gaming PCs. But it doesn't always lead in our benchmarks. It would likely fare better with more heavily threaded titles tested. Otherwise, the Core i5-8600K seems to be a better value pairing for Nvidia's GeForce GTX 1080.

The GeForce GTX 1070 pairs well with either a Core i5-8600K or Ryzen 7 1800X, though the latter notably offers additional performance in other enthusiast-oriented desktop apps. We've also found that the Ryzen 7 1800X is a great chip for game streaming, so give it serious consideration for more taxing workloads. If you're looking to go Intel, aren't interested in overclocking, and want to save a few dollars, the Core i5-8400 complements a GeForce GTX 1070 well, too.

On the value scale, Core i3-8350K and Ryzen 5 1600X both pair well with a GeForce GTX 1060. If you're willing to do a bit of overclocking, though, a Ryzen 5 1600 looks even better.

Speaking of overclocking, most of the CPUs we tested can operate at even higher frequencies. AMD has a much friendlier policy here; all of its chips are unlocked. Meanwhile, Intel charges a premium for its unlocked K-series SKUs. If you want to know more about available headroom, we have plenty of test results with every processor overclocked in their individual reviews, so head to those for an idea of potential gains.

Of course, our findings would be all the more interesting if we could run out and build a well-balanced machine with affordable hardware. Here's hoping the availability of graphics cards improves sometime soon!

MORE: Best CPUs

MORE: All CPUs Content

ASRock X399 Taichi Motherboard Review

Jacob Terkelsen — Mon, 05 Feb 2018 15:00:00 +0000

Layout & Features

Our X399 coverage continues with ASRock's X399 Taichi, a line upon which we've previously bestowed high marks, although our X370 launch experience with it was a bit bumpy. Let's see if the X399 Taichi brings style, performance, and value, even on the already expensive X399 platform.

Specifications

ASRock's Taichi lineup aims to deliver high-end performance and features while balancing aesthetics and an aggressive price point. The external packaging is clean and elegant. Its use of black, white, and grey is a pleasant departure from the myriad of reds and blues on the store shelves.

The decision to use a box flap keeps front and back panels simple, with the option of opening the cover to learn more about the bells and whistles. Minimalism carries over to the contents: a software CD, setup guide, postcard, four SATA cables, M.2 retention screw, WiFi antennas, and SLI bridges to accommodate dual, triple, and quad GPU configurations. It would have been nice to receive some stickers or a padded IO backplate with a $300 board.

Layout

Despite the number of connectors and on-board device support provided by the X399 chipset, ASRock still delivers a distinct steampunk pattern to the PCB. The illuminated gear-styled heatsink over the chipset is a very nice touch. Our G.Skill Flare-X RAM’s heat spreaders complement the cogs well, and our circular water block ties it all together. Because it's very neutral with its use of monochromatic colors, this board also pairs well with any color scheme and with RGB-lit parts.

The Taichi's USB 3.0 ports act as partitions between the board’s back I/O. The BIOS flashback button, a single PS/2 port, and a pair of USB ports are located directly above the two SMA headers for the built-in WiFi module. Two additional USB ports greet the digital SPDIF and five analog ports. The remaining USB 3.0 and USB 3.1 Gen 2 ports sit alongside the board’s two Intel-based gigabit Ethernet ports. This layout feels to be a more standard rear I/O than the MSI X399 board we previously reviewed.

ASRock’s Purity Sound makes an appearance, and a matte black plastic shell covers the EMI shielded components. The bottom segment of the board houses most of the headers and two HD audio ports, one of which is angled 90°. Two USB 2.0 headers also reside on the bottom section of the board right next to the clear BIOS button and numeric debug LED. A set of power and reset buttons sit directly above the first of two USB 3.0 headers, and the front panel header labels only the positive sides of the terminals on the silk screen.

The X399 Taichi gets storage right by providing eight angled SATA3 ports. You can deploy the U.2 port with ASRock’s U.2 kit or devices using the SFF-8639 port. If builders go the U.2 route, the M2_1 connector is disabled. Speaking of M.2, there are no fancy heat spreaders and two of the devices are placed between the four PCIe Gen3 x16 slots. The third M.2 option is placed below the chipset, but populating that space detracts from valuable planar graphics.

We were happy to see most power plugs on the right side of the board. An angled 6-pin PCIe header is provided for multi-GPU solutions while the 24-pin ATX header resides above the second USB 3.0 header. A duo of 12V CPU power headers—one 8-pin and one 4-pin—grant extra flexibility for power supply and cabling options.

Two fan headers are on the top side of the board, two are directly below the TR4 socket, and one is on the bottom sector. All are 4-pin. Three of these support 1A (12W) loads, while the CPU Optional and Chassis/Pump header support 1.5A (18W). As for RGB, there is a connection on the right and one on the bottom, both of which support 36W LED strips.

Threadripper’s wealth of PCIe connectivity is provided to the Taichi through four equally spaced Gen3 x16 slots routed out for x16, x8, x16, and x8 from top to bottom. A single open-ended PCIe Gen2 x1 slot rests directly in the middle of the PCIe complex. That bottom slot unfortunately rides up against all the bottom ports, so just make sure to use standard length, single-slot cards in that position.

The socket used for this review sample is the Lotes variety. We have not encountered any retention issues with this TR4 socket. We believe the layout of this board is just right, and given the inclusion of angled SATA and the power header placement, cable management is a breeze with the right case. For dual-card GPU configurations, there's plenty of space for ventilation if you're using smaller cards in the second and fourth slots.

MORE: Best Motherboards

MORE: All Motherboard Content

Software & Firmware

The Taichi’s UEFI is very stylish, sharing the product’s box art for its background. Luckily, the white text on black background works, but sometimes the longer menus get a little busy towards the bottom left of the screen. The Advanced Mode is our preferred viewing mode and we use OC Tweaker, Advanced, Tool, and the Exit screen the most.

The OC Tweaker panel for this product is consistent with other ASRock offerings intended for Ryzen-based platforms. You can tweak the APU/PCIE reference clocks for additional manual tuning. Manual mode lets us key in our desired frequency speed and target voltage for our processor. SMT mode is configurable, and below that are the XMP profile and memory configuration menus and submenus.

Below the memory options are the voltage and load-line settings. In Voltage OC Mode you can define whether to use fixed or offset voltage. Fixed voltage values can be adjusted in 6.25mV steps up to an alarming 2.5V for Vcore. As for offsets, a modest 0.300V is the maximum you can apply above the default VID for a processor. Load-line calibrations are defined in a numeric fashion, where Level 5 produces the most voltage droop and Level 1 gives advanced users a more constant voltage at the cost of efficiency.

The Advanced menu is fairly similar to AM4 platform options and gives lower level controls of the Ryzen CPU and chipset. A unique feature ASRock provides is an OC Mode (not to be confused with the previous tab) that enables three predefined voltage, frequency, and core configurations for a simpler overclocking experience. Do be cautious, because ASRock lists in all caps an articulate warning that these voltages could potentially damage your processor. We would recommend leaving the voltages alone and bumping the multiplier manually. More on that later.

Buried deep within the Advanced menu is the option to enable Threadripper’s highly differentiating feature, NVMe RAID. Otherwise, standard AMD PBS, CBS, and North/South bridge chip options are available for debug and stability improvements. If the advanced UEFI is just too much to consume, simply change the default Setup Style to Easy Mode. There are still “no help string” options sprinkled throughout the menus.

Within the tool menu, RGB lighting can be controlled through the RGB LED menu, and the standard color cycling modes are synchronized by the motherboard. We were also successful using the tool that let us flash our UEFI upgrade over the internet.

The hardware monitor section opens the controls for configuring the various fan headers to either 3-or-4 pin mode, PWM, and even switching between a standard header and a water pump header. We recommend switching the fan temp sources to something other than Tctrl due to the extremely elevated Tctl on Threadripper (do as we say, not as we do).

Up to this point the UEFI is well executed, but that FAN-Tastic tuning feature still murders our eyes with glaring blue that clashes completely with the Taichi theme.

MORE: Best Motherboards

MORE: All Motherboard Content

How We Test

Introducing The 570X (again)

The MSI X399 Carbon used our Thermaltake F51 Suppressor case, but ASRock's X399 Taichi gets to live its review life cycle inside the Corsair 570X, our new X399 test chassis. This case screams sex appeal, with tempered glass on four sides and three prominent 120mm LED fans in front with room to fit either 360mm or 240mm radiators for liquid cooling.

Since it is a review rig, we’ll be moving components in and out quite a bit, so the rubber pads for the glass standoffs have started to erode. Dust isn’t too much of a concern for us, but there are plenty of gaps throughout the glass joints, so we’ll keep a can of compressed air handy. As lovely as this case is, it is a chore keeping my kids from streaking their greasy little hands all over those pretty blue lights.

Another new addition to the bench is the Elgato HD60 Pro. We’re not currently testing it right now, but we intended to have some streaming benchmarks (and of course, hours of streaming with PlayerUnknown’s Battlegrounds). Plus, it gives us another reason to put cool hardware into a PCIe slot.

As for the remaining components, no other changes to our rig. Antec’s 1200W High Current power supply provides the juice to our Gigabyte GTX 1080 G1 Gaming GPU and Ryzen Threadripper 1950X. G.Skill provides the 32GB of DDR4-3200MHz RAM, and Toshiba still supports all of our motherboard reviewers with its 256GB RD400 M.2 NVMe drive. Fractal Design’s Celcius S36 360mm water cooler is placed into the intake side of our chassis to provide cooling for our 190W processor.

Comparison Products

ASRock X399 Taichi

MSI X399 Gaming Pro Carbon AC

The Hardware

Test System Configuration
Sound	Integrated HD Audio
Network	Integrated Gigabit Networking
Software
Graphics	NVIDIA
Chipset	AMD X370

We're sticking with our LG 4K display. 1080p data is useful in some instances, but it'd be wasteful paired with all of this sweet hardware. We will provide 1080p and 4K data for gaming benchmarks, and we are using 4K as our standard desktop resolution unless otherwise noted. If requested, we can provide 1440p as well in future articles.

Benchmark Settings

Synthetic Benchmarks and Settings
PCMark 8	Version 2.7.613 Home, Creative, Work, Storage, Applications (Adobe & Microsoft)
SiSoftware Sandra	Version 2016.03.22.21 CPU Arithmetic, Multimedia, Cryptography Memory Bandwidth
Iometer	Version 1.1.0 4k Random Read, 4k Random Write 128k Sequential Read, 128k Sequential Write
Cinebench R15	Build RC83328DEMO OpenGL Benchmark
CompuBench	Version 1.5.8 Face Detection, Optical Flow, Ocean Surface, Ray Tracing
3D Tests and Settings
3DMark 13	Version 4.47.597.0 Test Set 1: Skydiver, 1920x1080, Default Preset Test Set 2: Firestrike, 1920x1080, Default Preset Test Set 3: Firestrike Extreme, 2560x1440 Default Preset
Application Tests and Settings
HandBrake CLI	Version: 0.9.9 Sintel Open Movie Project 4.19 GB 4k mkv to x265 mp4
LAME MP3	Version 3.98.3 Mixed 271MB WAV to mp3 Command: -b 160 --nores (160 Kb/s)
Adobe After Effects CC	Release 2015.3.0 Version 13.8.0.144 PCMark driven routine
Adobe Photoshop CC	Release 2015.5.0 20160603.r.88 x64 PCMark driven routine (light and heavy)
Adobe InDesign CC	Release 2015.4 Build 11.4.0.90 x64 PCMark driven routine
Adobe Illustrator	Release 2015.3.0 Version 20.0.0 (64-bit) PCMark driven routine
Blender	Version 2.68a BMW 27 CPU Render Benchmark BMW 27 GPU Render Benchmark
7-Zip	Version 16.02 THG-Workload (7.6 GB) to .7z, command line switches "a -t7z -r -m0=LZMA2 -mx=9"
Game Tests and Settings
Ashes of Singularity: Escalation	Version 2.50.28527 High Preset - 1920x1080, Mid Shadow Quality, 1x MSAA Crazy Preset - 1920x1090, High Shadow Quality, 2x MSAA High Preset - 3840x2160, Mid Shadow Quality, 1x MSAA Crazy Preset - 3840x2160, High Shadow Quality, 2x MSAA
F1 2015	2015 Season, Abu Dhabi Track, Rain 1920x1080 - UltraHigh Preset, 16x AF 3840x2160 - UltraHigh Preset, 16x AF
Metro Last Light Redux	Version 3.00 x64 High Quality, 1920x1080, High Tesselation Very High Quality, 1920x1080, Very High Tesselation High Quality, 3840x2160, High Tesselation Very High Quality, 3840x2160, Very High Tesselation
The Talos Principle	Version 301136 1920x1080 - High Preset, High Quality, High Tesselation, 4x AF 1920x1080 - Ultra Preset, VeryHigh Quality, VeryHigh Tesselation, 16x AF 3840x2160 - High Preset, High Quality, High Tesselation, 4x AF 3840x2160 - Ultra Preset, VeryHigh Quality, VeryHigh Tesselation, 16x AF

We do have to make some changes and tweaks to our normal motherboard testing methodology for the X399 platform. For starters, some of the software in our suite does not support the plethora of cores available. As such, we have to deactivate resources in some instances and bypass certain features in others. These changes will be listed in the benchmarking results section, but we will be consistent from a test methodology perspective.

One Threadripper-specific feature that we covered in Paul Alcorn's CPU reviews is the use of Game Mode versus Creator mode. If I’m buying X399, I want all of my resources, regardless of the gaming penalties. Both modes will be tested in the benchmarks for the sake of this comparison, but moving forward we will only test with Creator Mode enabled.

MORE: Best Motherboards

MORE: All Motherboard Content

Benchmark Results & Final Analysis

Synthetics & Applications

In PCMark, the X399 Taichi paves the way with a 2.6% lead over the MSI and only lags in the Home benchmark.

Sandra follows suit with the Taichi scoring a win in nearly all categories. The Gaming Mode performance fails to keep up with Threadripper’s Creator Mode's use of cores and threads.

In our IOmeter tests, the Taichi's reads and writes suffer by small margins across the board. Cinebench R15 delivers the Taichi a slight defeat in single-threaded performance, but all 32 threads engage for a multi-threaded win. Compubench shows Creative Mode data sets within 0.1% of each other, but here, Gaming Mode is better for the Facedetect and Video Processing workloads.

Gaming synthetics start to favor the MSI motherboard slightly during the Skydiver benchmark, but Firestrike Extreme favors the ASRock X399 Taichi, and, interestingly enough, its physics score is ahead by 2.5%. With so much power, we're tempted to start running Time Spy and its variants against this powerhouse.

There isn’t much we can say about the application data this time. Clearly, more cores and threads favor the Creative Mode performance by wide margins in most of the workloads. The only notable benchmark is After Effects, where the MSI motherboard shows an impressive five second win. But for a platform with such a hefty price tag, we would have hoped to see a little bit better performance all around.

Test Results: Gaming & Performance

Ashes of the Singularity favors the Creative Mode’s core counts and shows double digit leads for both products when compared to Gaming Mode. The game gives a slight edge to the MSI X399 Gaming Pro Carbon AC at 1080p, and 4K shows similar trends with only a fractional win at the Crazy preset at normal batch levels.

F1 2015 is a virtual tie at 1080p, and 4K detail levels manage to let the MSI board squeak out a 1-3 frame win. Things aren’t looking good for the ASRock X399 in gaming benchmarks so far.

Metro Last Light Redux is a nice change of pace for the Taichi. We see wins at 1080p’s high and very high settings. Note, we do not factor in the maximum or minimums into performance calculations, because the workload is very noisy and we must filter out the extreme outliers (though our eyes do give us a good enough indicator for use in graphs). 4K resolutions see a similar trend with the high preset, but the very high preset equalizes all tested products and modes.

As we mentioned in the MSI X399 review, The Talos Principle unearths an interesting observation: as we increase the detail level, the scores tighten up, regardless of the resolution. We see this with the X399 Taichi's 5% lead over the MSI at 1080p high setting, but the 1080p ultra setting brings the difference down to 0.6%. Switching to 4K, the high setting is closer, with a 4% win for the Taichi, but goes back to a 0.6% win using the ultra setting. Game Mode is clearly the victor at 4K settings and all play modes reach frame rates in excess of 120Hz.

It comes as no surprise that Creative Mode data shows substantial wins in pretty much every category within our motherboard benchmark suite. Synthetic benchmarks and application data show that both the ASRock X399 Taichi and the MSI Gaming Pro Carbon Creative Modes are nearly 33% ahead of the Ryzen 1800X-esque Gaming Mode. However, our tested games still do not show the need for additional cores or processing threads.

But using the Gaming Mode as a baseline for comparison, the ASRock X399 Taichi wins in all averaged comparisons, even if by the slightest margins. Onto the physical tests.

Temperatures, Power & Efficiency

The Corsair 570X’s radiator placement clearly benefits the ASRock X399 Taichi, as it draws in fresh ambient air to cool the radiator (rather than the case air for exhaust). Even when only running Prime95 at stock settings, that temperature delta translates into roughly 6° cooler temperatures for the CPU die and Tctl sensor. We installed a 120mm fan above the CPU in an attempt to replicate the airflow of the Suppressor F51 exhaust going out the top of the case. The 570X is simply providing more air flow into the case than the single 140mm fan on the previous chassis. Regardless, both products show respectable temperatures and regulator heatsink designs that are capable of overclocking.

Threadripper is by no means a low power chip, but at idle the ASRock X399 Taichi is capable of drawing up to 25W less than the MSI X399 Gaming Pro Carbon, which is similar to its fellow Zen cousins. At all three remaining worst case load conditions, both boards perform within a few watts of each other and average wattage power favors the X399 Taichi.

With results very close in terms of average performance, efficiency boils down to the power results above. The ASRock X399 Taichi’s ability to draw nearly 30W less power at idle is enough to swing the win in its direction. It is good to see that ASRock’s boards continue to show excellent efficiency when compared to products in a similar price bracket.

Overclocking

The ASRock X399 Taichi utilizes an 11-phase regulator design that is split up as eight phases for the CPU Voltage rail and three for the SoC rail. The phases are routed to the left of the DIMM slots and above the socket so as to spread out the heat from critical spots in the board and increase surface area. As for the Vreg heatsinks, they are attached to each other by a heat pipe and contacted with the regulators with thermal pads. As we saw in the temperature charts, this design pays off, and under load the Vregs run slightly cooler compared to the contender.

Knowing our thermal and voltage limits from our previous test, we followed a similar overclocking procedure for the ASRock X399 Taichi. The 38 multiplier works right out of the box, and all 16 cores are able to run 12% above the advertised base clock speed. Beginning to drive into the 3900 MHz range requires adjusting the loadline to Level 3 and Level 2 to help reduce voltage droops caused by the starting and stopping of Prime95. With temperatures still in check, increasing voltage up to 1.25625V gets the Taichi to match the MSI’s result of 3975 MHz. Additional voltage doesn’t provide long term stability at 4.25 GHz, and temperatures begin to creep past the 86°C Tdie threshold to keep us from running 4.0 GHz this time.

Nevertheless, our 1950X sample is able to run at a lower voltage and similar temperature conditions while stably running at 3975 MHz on all cores while running Prime95 for eight hours. Our confidence improves by re-launching our benchmark suite to confirm our overclock under start-and-stop conditions. We also check our overclocks using the Ryzen Master tool, and it’s looking like AMD’s software is starting to mature a bit.

Memory overclocking is as simple as before, and DDR4-3333 is easily achievable after selecting the XMP profile and the enhanced frequency. Unfortunately, AGESA 1.0.0.4 did not improve our results in the latest tested beta UEFI, but we are confident that the Taichi can run faster if it's using DDR4-3600 or faster DRAM.

Value & Conclusion

Both competing boards are priced on the lower end of the X399 platform spectrum. For this high-end ecosystem, the ASRock X399 Taichi is listed at an MSRP of $349.99 and the MSI is $359.99. Sales and rebates can swing prices into the $330 range for either product at various merchants. Both boards provide similar capabilities, like WiFi, a storage and PCIe expansion, and high-end overclocking features.

On the one hand, the ASRock board wins the numbers game for both performance and price. However, the contents of the Taichi box are spartan, and the lack of M.2 heat spreaders and a removable WiFi module could sway the value to the MSI X399 board. Both products encounter similar overclocking experiences and both UEFIs had their own issues—but none of it really detracts from usability or enjoyment.

The layout of the Taichi is more ideal for the configurations I like, and I prefer the theme and aesthetic of the steampunk black and white board. Given the performance, efficiency, and overall experience we'd certainly recommend a strong look at the Taichi.

MORE: Best Motherboards

MORE: All Motherboard Content

Gaming And Streaming: Which CPU Is Best For Both?

palcorn@outlook.com (Paul Alcorn) — Wed, 03 Jan 2018 14:00:00 +0000

Introduction

Game streaming is immensely popular right now. And with more than 2 million broadcasters serving approximately 100 million viewers a month, Twitch is unquestionably the leading platform for gamers looking to show off their skills.

Streaming is pretty easy to set up once you decide what you'll use to encode your content. There are several options, each with its own strengths and weaknesses. Hardware-accelerated encoders, powered by Nvidia's NVEnc, AMD's Video Coding Engine, and Intel's QuickSync, all promise to wrap your gaming experience up in a good-looking and easy-to-stream package without hammering your CPU. But it's generally agreed that these encoders sacrifice quality and flexibility compared to software-based alternatives.

Conversely, software encoding with the x264 library is easy alongside utilities like Open Broadcaster Software (OBS) and XSplit. It offers better stream quality than GPU-driven approaches, but at similar bit rates. That's important for services with defined bit rate limitations (like Twitch). Streaming can also chip away at bandwidth caps. Just one hour of streaming at 10 Mb/s equals 4.5GB of data, so you want the most efficient encoder possible.

A host processing-based approach offers more flexible configuration options. Better quality comes at a price, though. Software encoding is a CPU-intensive process, which obviously tugs the other way against our usual aim of running games at the highest frame rates possible.

Hitching, stuttering, lag, and generally bad performance curse gamers who try doing too much with limited software encoding resources. Thus, it's often recommended to use a secondary system for high-quality streaming. In fact, that's the best way to achieve an unencumbered gaming experience while also providing your audience a crisp video stream. That requires attaching your gaming rig to another computer via capture card or local area network to offload the encoding workload, though. Many casual gamers simply don't have the money for such a configuration.

Software encoding can have a huge impact on gaming performance. Until recently, if you wanted to stream and game using the highest-quality solutions on one PC, you almost assuredly had to buy a pricey CPU with lots of threads to throw at the problem. But now those high-end desktop processors are more accessible than ever. AMD's Ryzen 7 chips lowered the bar for getting lots of cores into affordable platforms, and most games can't fully utilize them. That leaves lots of horsepower leftover for streaming at a reasonable price.

Intel's Coffee Lake processors recently debuted with more cores as well, giving those Ryzen CPUs a run for their money in multi-tasked, multi-threaded workloads. Good thing, too. As we'll see, the previous-gen Kaby Lake architecture simply cannot compete against like-priced AMD alternatives. Today's story compares a number of different host processors in combined gaming/streaming workloads. Along the way, we developed test methodology you'll see used in our future CPU reviews.

MORE: Best CPUs

MORE: All CPUs Content

Test Setup

How We Tested

Repeatability is one of the most important components of any useful benchmark methodology. All tests have some degree of uncertainty, but we're looking for a minimal and consistent amount of variability. Results plagued by wild swings in performance from one run to the next aren't usable as accurate benchmarks.

As an example, we've yet to develop any reliable multi-tasking benchmarks. In response to reader requests, we have worked diligently to create a series of tests that measure gaming performance with background applications like Web browsers, email clients, media players, Discord, and Skype open. Windows' prioritization appears to be based on fickle and unexplained factors. The operating system suspends various background processes unpredictably during one scripted sequence, then leaves them fully active during the next (even when the test environment hasn't changed). This unpredictability becomes more, well, unpredictable, as the number of open applications increases. Switching Windows into Game Mode only complicates matters further. So far, we have no solution. Our multi-tasking experiments yield deltas from 5 to 15 FPS between successive runs, which means they land nowhere near our expectations for a reliable benchmark.

Luckily, game streaming is much easier to control. Encoding is a CPU-intensive task that chews up plenty of cycles, so Windows doesn't suspend or otherwise interfere with it. This allows us to create repeatable benchmarks without extreme outliers.

What We're Measuring

Evaluating game streaming performance works across two axes: game quality and stream quality. Of course, we'll measure average, minimum, and 99th percentile frame rates with and without streaming in the background. We'll also include our usual frame time and variance results, which become more important once we start streaming.

We also need to account for stream quality. That means recording the percentage of frames encoded. Each processor pushes different frame rates, so each run correspondingly generates a different number of frames. As such, we measure the percentage of frames successfully encoded as "% of Frames Delivered." In the test below, a Threadripper 1950X CPU encoded 98.9% of the frames generated by our gaming session, meaning it skipped 1.1% of the frames due to encoding lag.

We're streaming at 60 FPS, so we also measure stream quality by listing the percentage of frames encoded within the desirable 16.667ms (60 FPS) threshold. We also include the percentage of frames that land above and below the 60 FPS threshold, which helps quantify the hitching and stuttering a viewer would see on the stream. Subjective visual measurements are still important, so we'll call out tests that generate a bad-looking stream.

Open Broadcaster System

There are several software encoding applications, but we chose Open Broadcasting System (OBS) due to its flexible tuning options, detailed output logs, and broad compatibility with streaming services. We're using the x264 software encoder, along with YouTube Gaming for our streaming service. Any run that reports frames dropped due to networking interference is discarded.

Our ultimate goal is to develop a test that measures CPU performance, so we select parameters that remove the most obvious bottlenecks. Gaming at 1920x1080 with an EVGA GeForce GTX 1080 FE side-steps a GPU limitation (as much as possible). Encoding overhead isn't as high with lesser video cards that generate fewer frames per second. We also test with a 10 Mb/s upload rate, though you can stream at 6 Mb/s or less. Our Internet connection would accommodate up to 35 Mb/s uploads. To vary game selection, we chose Grand Theft Auto V, Middle-earth: Shadow of War, and Battlefield 1 for our tests.

There are several other scenarios we could have added to increase the complexity of our testing, such as a simultaneous video stream from a webcam, recording the game to the host system, or streaming to multiple services at once. We went with just one service to reduce the number of variables...at least for now.

Finding the best streaming options requires some tuning for every game and hardware configuration. There is a delicate balance between game performance on the host system and stream quality for the remote viewer, so fine-tuning is needed to yield the best mix. We picked somewhat general settings that offered a good range of performance by our subjective measure. We also stuck with options that'd establish a level playing field for a wide range of test systems. Just be aware that there are plenty of knobs to turn, some of which could offer better performance than the ones we use (lowering the stream to 30 FPS, for instance, cuts encoding overhead significantly)

Tuning the encoding presets is one of the most direct ways to adjust streaming performance and quality for your system's capabilities. Slower encoding increases compression efficiency, which provides better output quality and reduces compression artifacts. OBS has 10 presets ranging from "ultrafast" (the lowest-quality setting with the least computational overhead) to "placebo" (offering the best streaming quality and consuming the most host processing resources). The placebo setting is aptly named; there is certainly a rapidly diminishing rate of return on stream quality after passing the "slower" preset (two ticks before placebo). More strenuous settings can quickly cripple even powerful processors, particularly if you are streaming from a single host system. Placebo with care.

We split our test groups into three different classes. After evaluating a few Core i3- and Ryzen 3-class processors and determining that they can't stream effectively at our settings, we chose Ryzen 5 and Core i5 models for our entry-level systems. We used the "veryfast" encoding setting for this class of CPU. Naturally, higher-end processors, such as our Ryzen 7/Core i7 and Threadripper/Core i9 chips, offer more performance, so we use the "faster" and "fast" settings, respectively, for brawnier CPUs.

Because we're testing with different encoding presets, you cannot compare test results for the different classes directly.

Test System & Configuration
Hardware	Intel LGA 1151 (Z370)Intel Core i5-8600K, Core i7-8700KMSI Z370 Gaming Pro Carbon AC2x 8GB G.Skill RipJaws V DDR4-3200 @ 2666 and 3200 MT/sAMD Socket AM4AMD Ryzen 5 1600X, Ryzen 7 1800XMSI Z370 Xpower Gaming Titanium2x 8GB G.Skill RipJaws V DDR4-3200 @ 2667 and 3200 MT/sIntel LGA 1151 (Z270)Intel Core i5-7600K, Core i7-7700K MSI Z270 Gaming M72x 8GB G.Skill RipJaws V DDR4-3200 @ 2666 and 3200 MT/sAMD Socket SP3 (TR4)AMD Ryzen Threadripper 1950XAsus X399 ROG Zenith Extreme4x 8GB G.Skill Ripjaws V DDR4-3200 @ 2666 and 3200 MT/sIntel LGA 2066Intel Core i9-7900X, Core i9-7980XEMSI X299 Gaming Pro Carbon AC4x 8GB G.Skill Ripjaws V DDR4-3200 @ 2666 and 3200 MT/sAll EVGA GeForce GTX 1080 FE 1TB Samsung PM863 SilverStone ST1500-TI, 1500W Windows 10 Creators Update Version 1703Corsair H115i

Test System & Configuration

Hardware

Intel LGA 1151 (Z370)Intel Core i5-8600K, Core i7-8700KMSI Z370 Gaming Pro Carbon AC2x 8GB G.Skill RipJaws V DDR4-3200 @ 2666 and 3200 MT/sAMD Socket AM4AMD Ryzen 5 1600X, Ryzen 7 1800XMSI Z370 Xpower Gaming Titanium2x 8GB G.Skill RipJaws V DDR4-3200 @ 2667 and 3200 MT/sIntel LGA 1151 (Z270)Intel Core i5-7600K, Core i7-7700K MSI Z270 Gaming M72x 8GB G.Skill RipJaws V DDR4-3200 @ 2666 and 3200 MT/sAMD Socket SP3 (TR4)AMD Ryzen Threadripper 1950XAsus X399 ROG Zenith Extreme4x 8GB G.Skill Ripjaws V DDR4-3200 @ 2666 and 3200 MT/sIntel LGA 2066Intel Core i9-7900X, Core i9-7980XEMSI X299 Gaming Pro Carbon AC4x 8GB G.Skill Ripjaws V DDR4-3200 @ 2666 and 3200 MT/sAll EVGA GeForce GTX 1080 FE 1TB Samsung PM863 SilverStone ST1500-TI, 1500W Windows 10 Creators Update Version 1703Corsair H115i

MORE: Best CPUs

MORE: All CPUs Content

Intel Core i5 & AMD Ryzen 5

We generated our stream using the "veryfast" encoder preset for this class of processors.

AMD Ryzen 5 1600X

Intel Core i5-8600K

Intel Core i5-7600K

Battlefield 1

First, we run a set of baseline tests to gauge performance without an active YouTube stream. Intel's Core i5-8600K fares best, but the Core i5-7600K and Ryzen 5 1600X aren't far behind.

The story changes when we add streaming to the mix. All three CPUs lose varying amounts of performance, with Ryzen 5 1600X ending up on top (though not by much).

We have two entries in our charts for the Kaby Lake-based Core i5-7600K: normal and high-priority. First, let's talk about the normal results. Intel's 4C/4T architecture hamstrings the Core i5-7600K, and it loses nearly half of its average frame rate. Unsurprisingly, this has an impact on in-game frame time variance and unevenness measurements. Those numbers don't tell the entire story, though. Battlefield 1 is basically unplayable due to hitching and stuttering during our stream.

The stream quality chart shows that Core i5-7600K only delivers 23.6% of its frames during the test, which means it drops an astounding 76.4% of the frames. This also results in a completely unwatchable stream. Flip over to the CPU utilization charts; they tell a profound story. In short, Core i5-7600K doesn't have any headroom available to handle the parallelized encoding workload throughout most of our benchmark.

The OBS software has several different settings to optimize performance, and adjusting process priority is a common tactic to improve streaming on Kaby Lake-based processors. So, we selected the high-priority setting (marked HP Stream in the charts). Effectively, this allows the encoder to steal cycles from the game engine, resulting in lower frame rates. But the encoder successfully processes 100% of the frames. Our Core i5-7600K fell to 51.6 FPS, and its 99th percentile measurements nosedived. On the flip side, less performance means fewer frames to encode per second, which in turn boosts encoding efficiency. Surprisingly, 91.05% of the frames landed within the desirable 16.667ms range.

We could always dial-back the encoder preset to accommodate Intel's lackluster -7600K, but similarly-priced processors in the test pool handle these quality settings adeptly. While streaming, Core i5-8600K provides nearly the same in-game average frame rate as AMD's Ryzen 5 1600X and it encodes 100% of the frames. The -8600K's streaming quality isn't as impressive as Ryzen's: 89.44% of frames land within the desirable range and ~10% of frames fall either above or below the threshold. In our opinion, the stream looked fine, but wasn't as smooth as the Ryzen 5 1600X's output.

A glance at the 99th percentile FPS chart tells us that Ryzen 5 1600X delivers the smoothest gaming experience during our streaming session. It also has plenty of reserves left in the tank, even during the streaming workload. In other words, you could probably increase the encoding preset to a higher level than the competing processors.

AMD's Ryzen CPUs don't overclock as well as Core processors, though. We tuned each model to see if the Intel's available headroom turned the tables. The x264 encoder uses AVX instructions extensively, so it serves as a nice stress test for overclocking stability. No doubt, Coffee Lake's AVX offset could come in handy, though it might also result in lower frequencies during streaming if you activate the feature (or leave it on Auto in the UEFI).

Overclocking nudges the Core i5-7600K ahead of Ryzen 5 1600X during our baseline gaming test, but it does little to rectify the issues encountered during our streaming workload. Four threads just can't hack it. Streaming suffers even after adjusting the priority status. The -7600K does serve up a higher percentage of frames with the overclocked settings and normal prioritization, but it still costs the chip game performance. Priority adjustments have little bearing on gaming smoothness: the -7600K's 99th percentile measurements are terrible with both settings.

A 4.9 GHz overclock does boost the Coffee Lake-based Core i5-8600K's 99th percentile measurements. However, it still trails Ryzen 5 1600X. The tuned -8600K is certainly competitive thanks to a compelling gaming experience, but it can't match the Ryzen 5 1600X's streaming performance. It's apparent the encoder appreciates lots of threads, so Ryzen's 12 are a big advantage.

The frame time chart summarizes in-game performance well. Kaby Lake-based processors scribble their way across with extreme variance. The Core i5-8600K is also easy to spot during our streaming workload. Meanwhile, Ryzen 5 1600X provides a much more consistent in-game experience.

Grand Theft Auto V

Grand Theft Auto V scales well with additional host processing resources and tends to favor Intel architectures. It isn't surprising to see the Core CPUs lead during our baseline tests. But we weren't expecting the Core i5-8600K to maintain its advantage over the competition while streaming.

The Core i5-8600K delivers 100% of its frames during the test, though only 89.71% fall into the perfect 60 FPS bucket. We see a nearly even split of 5% above and below that mark. Visually, the stream looks fine. The -8600K takes more of a hit in the 99th percentile measurements, though it still leads the rest of our contenders.

AMD's Ryzen 5 1600X starts out with the lowest average frame rate during our baseline and falls to 60.7 FPS during the stream. It doesn't suffer too badly in 99th percentile measurements. Meanwhile, the chip provides stellar streaming performance, dropping no frames and almost landing at a smooth 60 FPS.

The Core i5-7600K provides a decent gaming experience during normal streaming, churning out 79.2 FPS. That's because it really isn't encoding frames, though—it drops 84.9% of them. The stream is unwatchable by any measurement. Again, we can boost the -7600K's encoding performance to 100% by specifying high priority, but then the gaming experience is terrible. In fact, large portions of the scene simply do not render correctly.

Overclocking the Core i5-7600K doesn't help; it's still entirely saturated during the streaming test. But tuning does benefit Intel's Core i5-8600K, which leads the pack in both average and 99th percentile metrics. Its streaming quality is decent as well.

The Ryzen 5 1600X's performance improves after overclocking, too. It doesn't provide the absolute best frame rate while streaming, but it is perfectly playable. Also, Ryzen 5 1600X still offers the smoothest stream, if only just barely.

Middle-earth: Shadow of War

Intel's Core i5-7600K continues to disappoint. It serves up an average of 51.4 FPS while streaming (using the high priority setting) and encoding 100% of the frames. Unfortunately, game smoothness goes out the window again. The 15.9 FPS 99th percentile measurement tells us everything we need to know.

The Ryzen 5 1600X offers decent gaming performance while streaming, but that Coffee Lake-based -8600K continues leading.

AMD's Ryzen 5 1600X offers the best of both worlds, though Core i5-8600K is also very capable. Perhaps tuning the encode could get more performance from the Intel chip. Then again, you could say the same thing about Ryzen 5, which offers the best mix of streaming and frame rates in its class. It's obvious from these results that a Kaby Lake-based Core i5 owner needs to dial back streaming settings dramatically or turn to faster hardware.

MORE: Best CPUs

MORE: All CPUs Content

AMD Ryzen 7 & Intel Core i7

Paying more for a processor should also get you more performance, right? With that in mind, we're stepping up to the "faster" encoder preset for this CPU class.

AMD Ryzen 7 1800X

Intel Core i7-8700K

Intel Core i7-7700K (Kaby Lake)

Battlefield 1

The Core i5-7600K disappointed us in the previous round of tests, and Core i7-7700K picks up where the i5 left off. This chip just isn't well-equipped for streaming, despite the additional four threads enabled by Hyper-Threading. The -7700K dropped 94.3% of our frames using the normal settings, which comes as a side effect of excessive CPU utilization. We triggered the high priority setting to try rectifying this situation, but were presented the same bipolar behavior as Intel's Core i5-7600K. There just isn't a straightforward way around saturated CPU cores. You can either stream well or game well, but you can't do both with Kaby Lake-based processors and OBS streaming at 1920x1080/60FPS.

The Ryzen 7 1800X bears down with eight cores and 16 threads to provide solid streaming and gaming performance.

Core i7-8700K, which brings Coffee Lake up to a Hyper-Threaded six-core design, also performs well in this test. Overall, it offers the best gaming performance while streaming, 99th percentiles included. It even ekes past Ryzen 7 1800X with 100% of frames encoded to the stream. AMD's Ryzen 7 delivers 99.9% of the frames, and dropping 0.1% doesn't concern us. The stream is still smooth.

Overclocked, Ryzen 7 1800X delivers an ever-so-slightly better stream. But the Core i7-8700K leads in overall gaming performance. Of course, the 1800X's 117.8 FPS average, while streaming, is plenty impressive. Both processors prove up to this task.

Our encoding workload pegs the -7700K's cores at 100%, while the high priority settings help bring down CPU utilization. Simply, that's because the CPU has fewer frames to encode as a result of lower gaming performance.

Grand Theft Auto V

The Core i7-8700K takes a healthy lead during the baseline and streaming tests. It even provides more in-game performance while streaming than an unencumbered Ryzen 7 1800X.

Ryzen 7 1800X again achieves lower gaming performance, but a slightly superior stream than Intel's -8700K. The differences are pretty much imperceptible, though. Kaby Lake, on the other hand, simply can't compete.

Turning up the clocks benefits Coffee Lake greatly. Core i7-8700K offers the best overall performance, while Ryzen 7 1800X is competitive.

Middle-earth: Shadow of War

Coffee Lake continues to impress with strong in-game and streaming performance. The Ryzen 7 1800X does encode a higher percentage of frames than -8700K within the desirable 16.667ms range, but it also generates fewer frames. AMD's Ryzen does have more headroom to spare during our streaming workload though, so it'd likely accommodate more taxing encode presets.

We find that the stream frame time variance isn't very noticeable unless the 16.667ms percentage drops below a 90% threshold, such as what we see from Intel's Core i7-7700K. The -7700K does serve up better frame rates while it's streaming, even besting AMD's Ryzen 7 1800X. But that goes to show why we need to keep an eye on performance from every angle. Core i7-7700K drops 40% of its frames during this test! That makes any gaming performance advantage meaningless.

Ryzen 7 1800X musters a solid 75.4 FPS during the streaming test and provides a crisp stream to YouTube, as well.

Overclocking amplifies the trends we've been talking about. The same inherent strengths and weaknesses are still apparent. Ryzen 7 1800X and Core i7-8700K are both solid choices for streaming.

The Core i7-7700K demonstrates lower CPU utilization than usual during this test. Surprisingly, that doesn't equate to a usable combination of in-game and streaming performance, though, suggesting other architectural influences may be affecting our numbers.

MORE: Best CPUs

MORE: All CPUs Content

AMD Threadripper & Intel Skylake-X i9

You expect the best when you drop $1000+ on a CPU for gaming and streaming simultaneously. So, we shifted to the "fast" setting for this round of tests.

AMD Ryzen Threadripper 1950X

Intel Core i9-7900X

Intel Core i9-7980XE

Battlefield 1

AMD didn’t design Threadripper for "just" gaming, particularly at lower resolutions. But an intense streaming workload might expose more of the architecture's benefits.

Threadripper does have several configurable modes to tailor to its response to various tasks, as we outlined in our AMD Ryzen Threadripper 1950X Game Mode, Benchmarked article. Game Mode and Creator Mode both impact various titles differently, and we expect those same trends to carry over to our streaming benchmark. As such, we tested the Threadripper 1950X both ways.

The Threadripper 1950X pulls into range of Intel's Skylake-X models during our baseline tests. Creator Mode (CM) exposes all 32 of the 1950X's available threads, yielding more potential horsepower than Game Mode (GM). The 1950X's Creator Mode also makes it possible to encode 100% of the test run's frames, while Game Mode drops 1.6% of them. That isn't a huge sacrifice, but Game Mode also causes Threadripper to average 26 FPS fewer. That means it isn't encoding as many frames, either. We would have expected higher frame rates from a lighter encoding workload, but disabling half of the 1950X's threads in pursuit of higher game performance doesn't always work well when you're streaming, too.

The 18C/36T Core i9-7980XE doesn't perform as well as the 10C/20T Core i9-7900X during the streaming tests. Rather, the -7980XE stumbles in the smoothness department as its 99th percentile frame rates fall below the Threadripper 1950X, despite a lead in average frame rates.

Aside from Threadripper 1950X in Game Mode, all of these processors encode 100% of the frames.

The Core i9-7980XE's game performance is higher than the Threadripper models during our streaming test after bumping up clock rates. It also delivers a much better 99th percentile measurement.

AMD's Threadripper configurations fare better after tuning, too. Both encode 99.9% of the frames we send their way, which is adequate for a quality stream.

Grand Theft Auto V

Grand Theft Auto V's in-game performance goes Intel's way during the baseline benchmarks. Moreover, Core i9-7980XE redeems itself when it comes time to stream. That chip does command quite a premium though, so its advantage doesn't necessarily represent the best value.

Again, Threadripper's Game Mode just doesn't appear to be ideal for streaming. Game Mode provides better baseline frame rates in this title, but it falls to the bottom of our chart once we start encoding video.

Middle-earth: Shadow of War

Once again, Intel's processors offer the best in-game and streaming performance. This isn't entirely surprising; Threadripper may even be overkill for this type of enthusiast workload. We'd expect the architecture to handle workstation-class production workflows more adeptly.

Core i9-7900X does encounter a hiccup as its 99th percentile scores fall below Threadripper 1950X's during the streaming workload. The 1950X in Creator Mode also experiences some variance, with 20.38% of its frames falling below the 16.667ms threshold.

Even after a bit of tuning, Core i9-7900X provides a higher average frame rate than the 1950X in Creator Mode during our streaming benchmark, but can't match AMD's 99th percentile performance.

The Threadripper 1950X's difficulty in Creator Mode while streaming Middle-earth is even more pronounced. Some aspect of OBS doesn't agree with Creator Mode and this one game. We ran the tests several times to ensure it was a repeatable phenomenon.

MORE: Best CPUs

MORE: All CPUs Content

Final Analysis

AMD democratized access to high core counts with an attractively-priced Ryzen portfolio. In a way, we can thank the company for Intel's newfound interest in competing on a $/core basis. Just look at the difference in our results from Kaby Lake to Coffee Lake. As a result, high-quality software encoding on a gaming PC is becoming more realistic to mainstream gamers.

Our testing is indicative of general performance trends. Given enough time and energy, you could almost certainly improve upon our results. Part of that is by design: we're using these settings to compare large groups of processors against one another on a level playing field, as opposed to wringing the most performance out of any one processor. While we can't use our benchmarks to make definitive statements about the possibilities with each chip, we can draw some fair conclusions about how certain architectures behave.

Encoding is a parallelizable workload. If software encoding is your primary goal, you'll definitely want to seek out CPUs with lots of cores and simultaneous multi-threading capabilities. Intel's quad-core Kaby Lake models illustrate how chips that once offered class-leading gaming performance can fall apart during streaming. You can boost their performance by using less intensive quality presets, lowering the streaming frame rate, or sacrificing some quality with GPU acceleration. However, competing processors offer much more performance than Kaby Lake at the same presets and roughly the same price.

Many streamers place video quality over maximizing the frame rate of whatever game they're playing, so your own priorities will largely dictate how you tune your system. In fact, turning on v-sync may be a good way to balance streaming and gaming performance.If you seek the highest in-game performance while you stream, Intel's Coffee Lake-based Core i7-8700K is a good fit. The Ryzen 7 1800X is also competitive and tends to offer better streaming performance. Using our settings, the 1800X also had more CPU headroom leftover for more taxing encode settings, if desired. Granted, some of that extra horsepower is due to the 1800X's lower gaming performance, which means there are fewer frames to encode.

Two extra cores on the Coffee Lake-based Core i5 certainly help its standing, but the lack of Hyper-Threading has a definite impact on streaming performance. In the end, a six-core Core i5-8600K is forced to battle the 12-thread Ryzen 5 1600X, which offers a more balanced profile. Overclocking does help Intel somewhat. It can't overcome the advantage AMD gets from a more thread-heavy architecture, but it shrinks the gap somewhat in streaming workloads.

If you're really serious about streaming and gaming at the same time, the highest-end desktop CPUs are an option. Just expect to pay dearly for them. Most enthusiasts are better served by mainstream processors. Intel's Core i9 models generally provide better performance than the Threadripper 1950X, but they cost more, too. The 1950X is a solid value choice that also offers a diverse range of capabilities.

There are plenty of other solid options for gaming/streaming, and this introductory round of tests only focused on high-end models from each family. We'll expand our testing to locked SKUs as we work through coming CPU reviews.

MORE: Best CPUs

MORE: All CPUs Content

MSI X399 Gaming Pro Carbon AC Motherboard Review

Jacob Terkelsen — Wed, 22 Nov 2017 14:00:00 +0000

Features & Specifications

We've been testing AMD's Threadripper and Socket TR4 over the past couple of months in between our X370 and B350 motherboard reviews featuring the Ryzen processor. We’ve come to realize the potential of the high-end desktop platform and its caveats, and it's time to expose what we've discovered in this next series of X399 motherboard reviews. And so we begin.

PCIe, cores, and memory capacity are Threadripper's big selling points. AMD touts this platform as ideal for streamers, content creators, and professionals looking to speed up and multitask workloads in high-end desktop PCs. To get there, AMD pieces together two Ryzen die onto a single chip to form a dual-chip module (DCM). This DCM enables multiple bricks of CCX (CPU Complex), Infinity Fabric, and of course gobs of direct attach PCIe and quad-channel memory. The X399 chipset handles remaining I/O, storage, and system needs.

Because Threadripper is a massive chip with an abundance of built-in I/O, X399 has the straightforward goal of filling in the platform gaps. USB 2.0, native USB 3.1 Gen1, native USB 3.1 Gen2, large quantities of SATA3, and PCIe Gen2 can all be harnessed by motherboard manufacturers to provide additional device connections. It will be very interesting to see how different vendors pick and choose these options.

Specifications

MSI X399: First Foray into the Core Counts

On paper, X399 and Threadripper represent a potent combo. Cut from the same cloth as MSI’s other Performance Gaming boards, the X399 Gaming Pro Carbon AC ships with all of the bells and whistles we expect from a top-tier board.

The contents of the box are abundant, and a bulk of the cabling is delivered in a drawstring bag rather than plastic. Typical box contents include a driver CD, motherboard documentation, quick install guide, and SATA stickers. Unique to this MSI X399: a dual-card, high-bandwidth SLI bridge; an RGB splitter extension cable; mounting screws; USB extension; case badge; OC fan mounting hardware; and multiple sets of exchangeable covers for the various name plates around the board.

One item deserves its own five minutes of fame: A break-out PCIe card that houses the Intel Dual Band Wireless-AC 8265 WiFi module and antenna headers. This solution occupies a lone PCIe x1 slot and a USB 2.0 header on the board. We'd recommend populating it in the bottom x1 slot. If you're using SLI or three dual-slot PCIe cards, you're forced to put the card into the bottom x16 slot, which is less than ideal. However, you also don't have to use it. And it can be used in any other PC you've got.

Don’t be fooled by the silver-lined PCIe slots, which aren't all Gen3 x16. The first and third slots are wired for true x16 connections, while the second and fourth slots are wired for x8. As such, you should plug SLI or Crossfire into PCI_E1 and PCI_E4. Beyond that, you get two x1 PCIe slots. Through some clever slot placement, creative builders can consider the potential for capture cards and additional high-performance network or storage controllers.

Another feather in the cap for X399 is its ability to expose PCIe lanes for up to three native NVMe Gen3 x4 solid state drives. These cards are housed beneath each of the top three x16 slots, and as an added bonus you get M.2 heat shield plates. We use the Toshiba RD400 in all of our test beds, so we can’t commit to any vendor’s cooling solution to verify the effectiveness of the shields, but they are removable using an interesting hinge mechanism.

After getting the more prominent features out of the way, let’s tour board placement. Gazing at the top of the backplate, BIOS flashback and clear CMOS buttons are readily available along with two legacy USB 2.0 and a solitary PS/2 port. Beneath that is the cluster of eight USB 3.1 Gen1 ports, followed by a gigabit Ethernet port and two USB 3.1 Gen2 ports of the Type-C and Type-A variety. MSI tops it off with 5-port analog audio and one SPDIF digital audio out. The Audio Boost solution on the Gaming Pro uses the Nahimic 2+ technology, which the company claims improves audio customization and delivers improved SNR and amplification.

It’s hard to dress up fan headers, but having them neatly spaced and labeled at the bottom of the board is a clear benefit. Front panel audio, RGB and Rainbow, three 4-pin fan, TPM, and USB 2.0 headers fill the left half of the lower edge of the board. The right half has a USB 3.1 Gen2 header for advanced front panel chassis connections, built-in power and reset buttons, debug LED, and the Game Boost knob. The knob isn’t too critical for us since we’re using a closed chassis, but luckily it is accessible through the UEFI and can be engaged in seven different OC states, which we’ll cover later.

We love the two angled USB 3.1 Gen1 headers on the board. Cases use very stiff USB 3.1 cables, so reducing flex on the header makes us feel better. SATA connections are available through six angled and two straight connectors that support the Ryzen standard RAID 0/1/10 modes. The top right corner is very tight but gives us access to a 4-pin pump and CPU fan header. Let’s not forget about the 4-pin fan header located to the left of the DIMM slots.

Given the ample board size, it’s amazing how cramped the top half of the MSI X399 Gaming Pro Carbon AC can feel with eight DDR4 DIMM slots, one 24-pin ATX connection, two 8-pin CPU power connections, and one 6-pin PCIe power header. Things are so cramped that MSI opted to use a more vertical solution for cooling the Vregs of the board rather than the typical arrangement next to the back-panel portion. As odd as it looks in the product shots, this solution is much better than the elaborate designs we’ve seen in the past. After adding the cooling solution and populating the DIMM slots, the odd heatsink looks right at home. If additional cooling for the Vregs is necessary, simply employ the OC fan mounting hardware to point a fan directly at the heatsink. Using a top exhaust fan worked just as well for our overclocking tests.

Finally, we want to discuss the socket for TR4. We’ve always felt comfortable with the AM4 solution, where you place the CPU into the socket and activate the retention arm. AMD has been using that for decades it seems. The introduction of the TR4 retention system is clearly a deviation from the norm, but its size shouldn't intimidate anyone. The design is proper and works well in theory. Although the ratcheting mechanism seems like overkill, that audible click of the torque wrench ensures proper tension on the screws. The marking of the screw order is very helpful, too, because you might just tighten or loosen them willy-nilly.

Our only qualm is with the Threadripper cassette. With all of these solid metal socket components, the flimsy plastic material that surrounds Threadripper just seems insufficient. Also, the sled that the processor slides into is a little undersized. We did have to double-check several times that our processor was in fact lined up with the sled before placing the module down for retention. There have been discussions about proper screw length, and about Foxconn vs Lotes-produced sockets. The product we reviewed uses the Lotes variety. And for goodness sake, don't put your entire body weight’s pressure on it hoping to make better contact (we’ve seen that movie, it doesn’t end well -ed).

The aesthetic of the MSI X399 Gaming Pro Carbon AC is a lot less gaudy than some of the AM4 boards on the market, and the option for users to replace the coverings over the X399 chip, Audio boost, and I/O port can help give a little bit more personality. On-board LEDs are not the focal point of this board and instead help pull the focus to installed components. The only layout changes we'd propose would be moving the PCIe power connector next to the 24-pin ATX connector, and shuffling the two vertical SATA ports elsewhere. Also, shuffling the M.2 connectors and PCIe slots around so that the bottom x16 slot doesn’t hang over the bottom pins and debug header would enable the use of longer cards in that bottom slot.

MORE: Best Motherboards

MORE: All Motherboard Content

Firmware: The MSI UEFI Tweaks

Having reviewed a few of the MSI AM4 motherboards, we are quite comfortable with both the EZ and Advanced modes inside the UEFI. Our MSI X370 Krait Gaming review contains a lot of our UEFI information, but we are going to cover most of the deltas from other products here.

EZ Mode is the default user interface and gives us a good understanding of the system health, configuration, and access to the boot order. This is also a good place to keep in mind our core counts, because if we're running in Gaming mode (more on that later) we should observe a lower core count than usual. (The UEFI we tested was version 1.52, which implements the AMD AGESA 1.0.0.3A code base.)

Jumping straight into the overclocking menus, an included performance regulator enables some specific tweaks to the firmware to enhance performance for various benchmarks. Out of the box, this is disabled, and we intentionally checked to make sure it was turned off. We notice that the downcore feature is available, however we are examining its implementation with the X399 platform. Given that on Ryzen the feature enables/disables cores at the CCX level, we need additional testing to see how it works as implemented with Threadripper.

We mentioned that the X399 Gaming Pro Carbon AC has MSI’s Game Boost knob feature, and within the overclocking menu we can switch it from BIOS to physical control. This can also be changed in the top left quadrant of the screen by hitting the center button and then selecting the correct numerical mode, if necessary. Luckily, the help menu for this feature is well documented, and settings seem realistic, unlike similar options from other manufacturers. A default overclock of 400 MHz is completely feasible with this 1950X, and other parts might be comfortable there also. Just be cautious of your voltage.

The Hardware Monitor menu through the UEFI is a familiar face and gives readouts for system voltage, fan speed, and limited temperature sensors. Our included picture (above) is observing the pump for our newest AIO solution so we can monitor and change our fan curves to be more aggressive or shallower for intake and exhaust fans.

A couple of lessons learned: By default, IOMMU is set to “Auto.” Our Threadripper system was unable to boot with this setting, so we suggest switching this to “disabled” unless the deployed application requires virtualization.

MORE: Best Motherboards

MORE: All Motherboard Content

How We Test

Because this is our first X399 motherboard review, we'll detail our prosumer rig component choices. First, we're rocking a Ryzen Threadripper 1950X, with its 16 cores and 32 threads; because it tops out at a TDP of 180W, we'll need proper cooling to reach maximum Turbo Core speed.

The Hardware

Test System Configuration
Sound	Integrated HD Audio
Network	Integrated Gigabit Networking
Software
Graphics	NVIDIA
Chipset	AMD X370

Fractal Design provided its Celsius S36 360mm all-in-one cooler, which has the standard round-style baseplate attached to the pump head. AMD did a good thing by providing the TR4 socket adapter with the Threadripper processor so that no custom adapters are required.

We purchased a spanking new Gigabyte GTX 1080 G1 Gaming card to properly test this system.

To pump out enough power and wire up enough ports on our test boards, we have migrated an Antec High Current Pro 1200W power supply from other projects. Given the number of available PCIe slots and power connectors on the board, having the PSU's extra dedicated 12V connectors is a necessity, and we hope to one day draw every available watt from this 80 Plus Gold product.

We’ve had great success with G.Skill’s memory on our AM4 test bench, and the company provided two sets of Flare-X 2x8GB DDR4-3200 DIMMs for this testing. As much as we want to populate all of the DRAM slots, with X399 we'd be looking at close to $900 in RAM at today’s prices.

Toshiba continues to provide our M.2 NVMe drive of choice, the RD400 256GB SSD. We'd love three of them, but Thomas believes that would be too greedy.

Unfortunately, our X399 chassis hasn't arrived yet, so we're borrowing the Thermaltake F51 Suppressor from our AM4 rig and will be installing the radiator at the top to exhaust air out of the case. We will reintroduce the Corsair 570X in our next review, placing the radiator in the front of the system in order to utilize all three of its gorgeous 120mm RGB fans as intakes across the radiator and into the system.

We're sticking with our LG 4K display. 1080p data is useful for some instances, but it'd be wasteful paired with all of this sweet hardware. We will provide 1080p and 4K data for gaming benchmarks, and we are using 4K as our standard desktop resolution unless otherwise noted. If requested, we can provide 1440p as well in future articles.

Benchmark Settings

Synthetic Benchmarks and Settings
PCMark 8	Version 2.7.613 Home, Creative, Work, Storage, Applications (Adobe & Microsoft)
SiSoftware Sandra	Version 2016.03.22.21 CPU Arithmetic, Multimedia, Cryptography Memory Bandwidth
Iometer	Version 1.1.0 4k Random Read, 4k Random Write 128k Sequential Read, 128k Sequential Write
Cinebench R15	Build RC83328DEMO OpenGL Benchmark
CompuBench	Version 1.5.8 Face Detection, Optical Flow, Ocean Surface, Ray Tracing
3D Tests and Settings
3DMark 13	Version 4.47.597.0 Test Set 1: Skydiver, 1920x1080, Default Preset Test Set 2: Firestrike, 1920x1080, Default Preset Test Set 3: Firestrike Extreme, 2560x1440 Default Preset
Application Tests and Settings
HandBrake CLI	Version: 0.9.9 Sintel Open Movie Project 4.19 GB 4k mkv to x265 mp4
LAME MP3	Version 3.98.3 Mixed 271MB WAV to mp3 Command: -b 160 --nores (160 Kb/s)
Adobe After Effects CC	Release 2015.3.0 Version 13.8.0.144 PCMark driven routine
Adobe Photoshop CC	Release 2015.5.0 20160603.r.88 x64 PCMark driven routine (light and heavy)
Adobe InDesign CC	Release 2015.4 Build 11.4.0.90 x64 PCMark driven routine
Adobe Illustrator	Release 2015.3.0 Version 20.0.0 (64-bit) PCMark driven routine
Blender	Version 2.68a BMW 27 CPU Render Benchmark BMW 27 GPU Render Benchmark
7-Zip	Version 16.02 THG-Workload (7.6 GB) to .7z, command line switches "a -t7z -r -m0=LZMA2 -mx=9"
Game Tests and Settings
Ashes of Singularity: Escalation	Version 2.50.28527 High Preset - 1920x1080, Mid Shadow Quality, 1x MSAA Crazy Preset - 1920x1090, High Shadow Quality, 2x MSAA High Preset - 3840x2160, Mid Shadow Quality, 1x MSAA Crazy Preset - 3840x2160, High Shadow Quality, 2x MSAA
F1 2015	2015 Season, Abu Dhabi Track, Rain 1920x1080 - UltraHigh Preset, 16x AF 3840x2160 - UltraHigh Preset, 16x AF
Metro Last Light Redux	Version 3.00 x64 High Quality, 1920x1080, High Tesselation Very High Quality, 1920x1080, Very High Tesselation High Quality, 3840x2160, High Tesselation Very High Quality, 3840x2160, Very High Tesselation
The Talos Principle	Version 301136 1920x1080 - High Preset, High Quality, High Tesselation, 4x AF 1920x1080 - Ultra Preset, VeryHigh Quality, VeryHigh Tesselation, 16x AF 3840x2160 - High Preset, High Quality, High Tesselation, 4x AF 3840x2160 - Ultra Preset, VeryHigh Quality, VeryHigh Tesselation, 16x AF

One Threadripper specific feature that we covered in Paul Alcorn's CPU reviews is the use of Game Mode versus Creator mode. If I’m buying X399, I want all of my resources, regardless of the gaming penalties. Both modes will be tested in the benchmarks for the sake of this comparison, but moving forward we will only test with Creator Mode enabled.

MORE: Best Motherboards

MORE: All Motherboard Content

Benchmark Results & Conclusion

For this first X399 review, we have to include some comparison data. We seeded a mysterious challenger that will be covered in our next review.

Synthetic Benchmarks

PCMark is our obligatory first benchmark to examine general purpose performance. We observe our first outdated benchmark in Creator Mode, in that its Creative suite fails to complete. The 4K H.264 application was built with 32-bit addressing and crashes with more than 20 cores enabled. Futuremark states that PCMark 10 fixes this issue, so we look forward to moving to that revision soon. In the Adobe suite, the MSI X399 Gaming Pro Carbon AC performs up to 3% slower in Creative Mode compared to Gaming Mode. This was a likely outcome, given that Adobe CS is lightly threaded.

Sandra synthetics scale almost perfectly with the arithmetic workloads across core counts but show the MSI X399 trailing the mysterious competitor by a few percentage points. Performance scaling starts to dwindle with the multimedia, crypto, and memory workloads, and the MSI still lags by a few percentage points.

Now that we’re using IOmeter to gauge our disk performance, we observe faster random reads and writes from the Gaming Mode for the MSI X399 board. With sequential reads and writes, the MSI X399 also grabs its first win of the review versus its masked contender.

No surprises with Cinebench. Gaming Mode uses half as many threads and sees roughly 46% less performance in the multi-threaded scenario. Less than one percentage point separates all the data sets in single-threaded mode, which is expected.

Compubench begins our shift in emphasis from the CPU to the GPU in some applications, and FaceDetect, video processing, BitCoin, and Fluid 64k metrics clearly favor the CPU latency reductions when in Gaming Mode. Creator Mode data sets are not in the MSI X399’s favor, suffering by 3% in FaceDetect.

Let’s shift gears to the 3D synthetics: If you’re running 720p on a Threadripper system, we feel sorry for you. However, we’ve got the data, and surprisingly the graphics and combined scores for the Creative Mode MSI X399 are clear wins, with the Gaming Mode results showing a significant, yet expected dip in physics results. Standard Firestrike appears to show comparable results for both Ryzen Master modes, but the generic X399 contender does inch ahead by a rounded 2% on all metrics. At 1440p we are starting to stretch into the real environment for our test bench, and the Gaming Mode’s physics test is what drags that configuration down.

Timed Applications

For applications we’re going to see a lot of the same results throughout, because with Creative Mode we have twice as many cores and twice as much RAM available. This translates into time reductions in Handbrake, both Blender tests, and 7-Zip, shaving up to three minutes off some workloads.

Unfortunately, Adobe is still lightly threaded and doesn’t enjoy the resources of Threadripper, and actually prefers the less endowed configuration. For some reason, Adobe Illustrator finishes almost one second faster on the MSI X399 motherboard.

Gaming Performance

We would be crazy if we recommended people not game on Threadripper rigs. Coupled with a GTX 1080, this platform is built for gaming and performing computational work at the same time. What will be interesting to see is how different games behave with the different Ryzen Master operating modes.

Ashes of the Singularity: Escalation starts us out at 1080p with comparable performance at both high and crazy presets for both Creative Mode contenders. Comparing the operating modes shows a performance advantage for the higher core count configs with the highest increase coming from High presets. The trend continues to a lesser extent as pixel count increases and the Crazy preset continues to squeeze all configs and systems below that sweet 60 FPS threshold.

F1 2015 throws us the only curve ball in the gaming segment; it doesn't want to use all 16 cores. By adjusting our test XMLs to the 8-core setting with hyperthreading enabled, the game is functional and utilizes all the address space available. The MSI X399 Gaming Pro Carbon AC enjoys a 30% performance increase at 1080p by doing so, while 4K levels the playing field. In the end, the MSI X399 takes the checkered flag with F1 2015.

Metro Last Light scaled very well in our AM4 coverage, and it comes as no surprise that the scaling continues when keeping product and core counts consistent. This is the first instance of the Gaming Mode setting drastically improving gaming performance, at least at 1080p. At 4K and during both graphical presets, all products are within two frames of each other.

Ugh, The Talos Principle. The MSI X399’s Creative Mode lags behind its Gaming Mode by about 10% at 1080p and High presets, while the Ultra presets are only lagging by about 6%. At 4K we see nearly 15% performance increases at the High preset, but Crazy levels everything back out across the platforms. Taking all data into account, this hardware is complete overkill for this game.

For the sake of simplicity, we are going to normalize our data around the Gaming Mode results for all of our samples today. By definition, the Gaming Mode Threadripper processor is, in effect, a Ryzen 7 processor, so we can extrapolate several key messages regarding performance. First, there are definite gains to be had with Threadripper in our benchmark suite. With more threads and memory available for computing, Creator Mode achieves an advantage of at least 20% across the board. Next, the MSI X399 Gaming Pro Carbon AC suffers between 1% and 2.5% behind our mysterious contender on average. Maybe the power, efficiency, and value arguments can redeem this board.

Power, Thermal & Efficiency Results

For simplicity, we only tested using the Creative Mode setting, because power gating and other processor architecture settings could make direct power comparisons more complicated. Gazing at the Kill-A-Watt, the MSI X399 Gaming Pro Carbon AC system draws 25 more watts at idle than our competitor and overall draws roughly 10W more than the contender on average, which is within four percentage points. We’d be interested to see the statistics of the power output on our systems, so if anyone can recommend a graphing utility that plugs into USB, let us know in the comments.

We have abandoned AIDA64 for temperatures, because we've had issues with software updates and the impact of them on reproducible results on our AM4 platform. The first picture above is for reference, comparing AIDA64 vs HWiNFO. On TR4 sockets, we’ll be using HWiNFO and will be reporting the Tdie and Tctl from that tool. As expected, the MSI X399 runs a modest 44°C above ambient and reports the expected +27°C Tctl metric. The contender uses a different chassis and intakes ambient air into the radiator for an impressive 6°C decrease. Unfortunately, the airflow across the regulator on the MSI X399 board is just not optimal or cool enough to match the contender’s performance, although 37°C above ambient is still within safe regions on this system.

Because the MSI X399’s overall performance metrics are close to those of the contender, the power draw data is what dictates the efficiency score for the Pro Carbon. We don't think 4% additional power is a terrible thing, and there could be some board to board variation that plays into either vendor’s favor.

Overclocking

Circling back to board components, the MSI X399 Gaming Pro Carbon AC utilizes a 10+3 phase design for power delivery directly to the Threadripper processor. It is hard to believe that the designers could fit that many CPU phases directly beneath the limited footprint of that tower of a heatsink. Fiddling with the power workloads for our testing, we encounter odd occurrences of the system just shutting off, which is indicative of a power delivery problem. After a few emails with the MSI technical staff, it turns out some of the preliminary settings with the UEFI were based on engineering samples provided by AMD and didn’t represent the entire population of processor vintages. Upgrading to the 152 UEFI remedies this problem as we get our testing underway.

Having a full 360mm radiator at our disposal is new ground for this reviewer and with that comes optimism that we can achieve better results than our dud of a Ryzen 7 1700X. I am also determined to best Paul’s standard overclocking efforts. I want to break 3950 MHz.

Bumping up the multiplier to 38 right out of the gate shows all 16 cores humming right along and, interestingly enough, temperatures not much higher than our standard testing. Bumping up to 3850 and then 3900 prompted us to increase the loadline to level 2, which is consistent with our X370 observations as well. With temperatures still within reason, increasing the voltage up to 1.2875 ultimately enables us to beat Paul and run at an impressive 3975 MHz! 4000 MHz is stable on this platform for limited stress testing, but 3975 MHz ran all night and well into the morning (I forgot to shut it off before work). Our temperatures during extreme testing were at 85°C Tdie, which is as high as this reviewer feels like pushing the system.

Given the cost of DDR4 right now, it makes sense to only populate each of the quad-channels with a single stick of RAM. If we reference our AM4 motherboard reviews, this type of configuration on the memory controllers for each CCX is optimal for overclocking with very little need for code enhancements. DDR4-3200 is our highest available XMP profile, and this system boots it right up without AMD’s advanced boot UEFI options. Even better, this DDR4 kit can also bump up to DDR4-3333 MHz without a single gander at the subtimings, voltages, or loadline settings.

Because AMD selectively bins out Ryzen die to populate the Threadripper lineup with lower voltages, it is safe to say that using the Ryzen 7 overclocking limits listed around the internet is not safe or practical for these MCMs. Keep temperatures within reason for the installed cooling solution and keep the voltages below 1.45V at the die for short-term wins. I don’t feel like burning an $800 hole in my pocket, so 1.2875V and 85°C Tdie will work just fine. Please follow AMD’s guidelines for overclocking.

Overclocking on this MSI X399 Gaming Pro Carbon AC is a breeze when using the latest UEFI. With the added CPU power connectors, it is clear that one must be prepared to pay the electric bill when running a Threadripper overclocked, as we are reporting wattage readouts upwards of 400W during Prime95 small FFT tests. Overall, this is great success for MSI.

Value, Conclusion & Teasers

We are going to use the Gaming Mode performance metrics of the MSI X399 Gaming Pro Carbon AC as a value baseline. Given the performance win and lower price of its competitor, the MSI X399 doesn’t win the numbers battle this time around. However, this board is still a worthy choice. The added Wi-Fi module is a great addition and worth probably $40 on its own.

Threadripper has proven to be a powerhouse compared to our X370/B350/Ryzen test bench. Unfortunately, where this system truly shines is in workloads that we just don’t heavily emphasize in our motherboard test suite. If you’re looking for more validation that Threadripper is for you, head on over to Paul’s Threadripper reviews, where he provides more analysis on the compute side and compares the chip against some of the Blue Team’s products.

MORE: AMD Ryzen Threadripper 1950X Review

MORE: AMD Ryzen Threadripper 1920X Review

MORE: AMD Ryzen Threadripper 1900X Review

As for TR4, it has options all around for I/O, expansion cards, USB, and storage that would cost an arm and a leg to fully populate. That loadout would also be unrealistic for a test bench, and our applications probably won’t support the RAM, power draw, or odd-ball drivers that might be required for true performance in a workstation system. Since every Threadripper provides the same I/O to the chipset, it is not as difficult to pick out a motherboard compared with other chipsets, where manufacturers can get squirrely with their wiring.

The MSI X399 Gaming Pro Carbon AC is a great choice for a reference system. Performance measurements are lower than our mysterious sample, but the layout, options, and the complete package line up perfectly with the definition of X399’s market. This board doesn’t pack a lot of flair, but we can appreciate the beauty of simplicity. Top it all off with a clean UEFI and an accessory kit that is fully functional, and this board earns the Tom’s Hardware Approved award.

MORE: Best Motherboards

MORE: All Motherboard Content

Intel Core i9-7960X Review: Skylake-X At 16 Cores

palcorn@outlook.com (Paul Alcorn) — Mon, 25 Sep 2017 09:20:00 +0000

Features & Specifications

It's certainly been a great year for enthusiasts. The dynamics of the desktop CPU market changed radically as AMD rose up to challenge Intel's dominance for the first time in recent history. For us, that meant an almost endless stream of new processors to benchmark and compare.

AMD's attack began with mainstream and entry-level parts offering more cores at lower price points than comparable Core CPUs. Intel continued on its merry way, introducing a first round of expensive Skylake-X CPUs that, in many of our benchmarks, failed to impress. When the Ryzen Threadripper family launched shortly thereafter, wielding up to 16 cores and the ability to schedule as many as 32 threads, it became clear that AMD meant business.

All along, we knew Intel was also preparing 12-, 14-, 16-, and 18-core processors of its own. But the company seemed to be rushing around in response to its perhaps unexpected competition. The Core i9-7900X is fast, no doubt. However, excessive thermal and power consumption, along with limited overclocking headroom, marred its debut. Although Intel took steps to drop prices, its Skylake-X models still sell for a rich premium compared to the Threadripper chips contending for enthusiast affection.

From what we understand, Intel believes it has an ace in the hole. Those 6-, 8-, and 10-core Skylake-X CPUs it already sells were a mere prelude to the big guns landing today. Moving forward, Threadripper has to contend with much more sophisticated Core i9s, led by an 18-core -7980XE.

So, Where's The Core i9-7980XE Review?

Unfortunately, our sample and test platform aren't working well together, demonstrating unexpected Turbo Boost frequencies at stock settings. Specifically, the Core i9-7980XE does not exceed 3.4 GHz on a single core. For now, the source of this issue remains unidentified, despite many combinations of components, operating system revisions, and motherboard firmware revisions. We suspect the problem traces back to our engineering sample CPU or motherboard BIOS. We are, of course, in contact with both Intel and MSI regarding our findings.

Interestingly, the 16C/32T Core i9-7960X, which goes up against Ryzen Threadripper 1950X's core count (albeit at a $1700 price point) does work correctly with our test platform. This allows us to post a review of that model with accurate results. For now, our test suite is abbreviated. But we'll update this space with our complete list of benchmarks in the days to come.

Specifications

Intel's Core i9-7960X is the company's second-from-the-top Skylake-X model, selling for $300 less than the flagship -7980XE, but with two fewer cores and 4.75MB less cache. That puts it at a similar core/thread count as AMD's Ryzen Threadripper 1950X (16C/32T). Despite its discount relative to -7980XE, though, Core i9-7960X bears a massive premium compared to the top Threadripper model ($1700 vs. $1000). While we're confident that Intel gets more done per clock cycle with its Skylake architecture, the fact that AMD gives you just as many cores at such a discount is compelling in a great many workloads.

Intel hopes to offset higher pricing with superior performance and features, such as its new mesh topology. This allows the construction of a single monolithic die, as opposed to AMD's multi-die module. Skylake-X is consequently a significant retooling of the Skylake architecture, as we've covered in-depth on multiple occasions. In some applications, this can cause performance to slide the wrong way. Most tests benefit from Intel's enhancements, though. Company representatives claim that future software optimizations may help ameliorate the few cases where Skylake-X suffers, similar to what AMD went through as Ryzen came online.

A realigned cache hierarchy promises improved application performance, while dual 256-bit FMAs work in parallel to support AVX-512. Intel's combination of greater cache throughput and AVX enablement facilitate up to 1 TFLOP compute performance from a single processor. That's a first on the desktop.

Core i5-7640X

Core i7-7740X

Core i7-7800X

Clock rates generally drop as core counts increase. Intel does, however, offset a low-sounding 2.8 GHz base frequency with significantly improved Turbo Boost bins. Core i9-7960X delivers a 4.2 GHz Turbo Boost 2.0 bin with two cores active, and up to 4.4 GHz with Turbo Boost Max 3.0, targeting lightly-threaded workloads across two "favored" cores. The feature is supported natively in the latest Windows 10 Creators Update. However, you still need Intel's driver if your motherboard firmware doesn't implement the feature correctly.

Active Cores	1-2	3-4	5-12	13-16
Intel Core i9-7960X (GHz) Turbo Boost 2.0	4.2	4.0	3.9	3.6

Intel's sophisticated Turbo Boost algorithms provide accelerated clock rates based upon the number of active cores. These higher frequencies even kick in when 16 cores are active, nudging Core i9-7960X as high as 3.6 GHz so long as power, thermals, and current fall below certain thresholds. While Intel only guarantees -7960X's base frequency, we found the Turbo Boost bins to be consistently aggressive during our testing.

Intel's Skylake-X series is repurposed data center silicon from the Xeon line-up, so the processors share the same LCC (Low core Count) and HCC (High Core Count) die that power up to 10- or 18-core processors, respectively. Previously, Core i9-7900X weighed in as Intel's largest LCC-based desktop processor, but Intel employs its HCC die for the chips launching today. That means the -7960X has a much larger die area for dissipating heat. We're hoping this improves its thermal performance compared to Core i9-7900X.

As we know, Skylake-X could certainly use some help in that department. The -7960X's 165W TDP is 25W higher than the -7900X we've already had problems with. And unfortunately, Intel is still using thermal paste between its die and heat-spreader. Even the beefiest custom loops won't be able to overcome this limitation, barring direct-die cooling. Notably, this isn't an issue AMD has with Threadripper thanks to its use of solder.

Similar to the other Skylake-X CPUs, the -7960X supports up to DDR4-2666 memory. It's also used with the same X299 Basin Falls chipset, which boasts improved DMI throughput and 30 HSIO (High Speed Input/Output) lanes that motherboard vendors can allocate to expanded connectivity options. Core i9-7960X also enables the full complement of Skylake-X's 44 PCIe 3.0 lanes, though it trails Threadripper. All of AMD's high-end desktop processors wield 64 lanes of third-gen PCIe. Of course, finding a use for that many lanes might be challenging, but it's certainly nice to have them available.

MORE: Best CPUs

MORE: All CPUs Content

Overclocking & Test Setup

Overclocking

If we learned one thing from out overclocking efforts, it's that Intel's thermal throttling mechanisms work great. And often. Even slight voltage increases result in large temperature increases. We found that adjusting the VCCIN voltage (voltage going to the processor) in tandem with VCORE adjustments (voltage to the cores) delivers the best results. It's even possible to generate higher multi-threaded performance scores in some benchmarks, like Cinebench, at the same frequency but with increased VCCIN voltage.

We raised VCCIN up to 1.9V, though this does cause thermal output to increase. We also found that up to 2.0V is "generally" safe if you are using water cooling. However, we lowered the VCORE to a mere 1.1V to combat heat. That gave us a semi-stable 4.5 GHz overclock; we were able to run a wide range of heavily threaded workloads, though extended AIDA stress tests exposed throttling. We decided to stick with a 4.3 GHz all-core overclock (84-88°C) because we found that to be the safest setting that wouldn't trigger the aggressive throttling algorithms. The -7960X is very sensitive to increased voltages, and even bumping up to 4.4 GHz resulted in nagging throttling during stress tests.

The processor will throttle if it overheats, but Skylake-X motherboards also throttle if the package power exceeds a defined threshold. With our MSI motherboard, that ceiling lands in the 300W range. We can remove the power restrictions via BIOS manipulation and pull more than 400W through the package, but we shied away from that because we want to test this processor in the future.

We also tested with both the AVX and AVX2 offsets at the Auto setting. The offset automatically reduced our AVX frequencies to 2.4 GHz. We didn't spend too much time tuning this option due to time constraints, but given the platform's high power consumption and thermal challenges, the dual-headed offsets are a welcome and needed feature.

Be sure to bring a custom loop if you plan to overclock. Also, we advise air or water cooling on the power delivery subsystem. Make sure to use a power supply able to deliver a minimum of 20A on the 12V rail. We've heard reports of up to 530W flowing through the eight-pin EPS cable, so a beefy PSU is a must.

Fine-grained tuning will yield better overclocks as we all acclimate to tuning Skylake-X. However, throttling happens any time you take a wrong turn with your settings. A lot of trial and error is required in order to get more from these components.

Comparison Products

AMD Ryzen Threadripper 1950X

AMD Ryzen Threadripper 1920X

Ryzen 7 1800X

Test Systems

Test System & Configuration
Hardware	Intel LGA 2066Intel Core i9-7960XMSI Xpower Gaming AC 4x 8GB G.Skill Ripjaws V DDR4-3200 @ 2666 and 3200 MT/sAMD Socket SP3 (TR4)AMD Ryzen Threadripper 1950X, 1920XAsus X399 ROG Zenith Extreme4x 8GB G.Skill Ripjaws V DDR4-3200 @ 2666 and 3200 MT/sIntel LGA 2066Intel Core i9-7900X, i7-7820XMSI X299 Gaming Pro Carbon AC4x 8GB G.Skill Ripjaws V DDR4-3200 @ 2666 and 3200 MT/sAMD Socket AM4 AMD Ryzen 7 1800XMSI X370 Xpower Gaming Titanium2x 8GB G.Skill Ripjaws V DDR4-3200 @ 3200 MT/sIntel LGA 1151 Intel Core i5-7700K MSI Z270 Gaming M72x 8GB G.Skill Ripjaws V DDR4-3200 @ 2666 and 3200 MT/sAll EVGA GeForce GTX 1080 FE 1TB Samsung PM863 SilverStone ST1500, 1500W Windows 10 Creators Update Version 1703Corsair H100i v2

Test System & Configuration

Hardware

Intel LGA 2066Intel Core i9-7960XMSI Xpower Gaming AC 4x 8GB G.Skill Ripjaws V DDR4-3200 @ 2666 and 3200 MT/sAMD Socket SP3 (TR4)AMD Ryzen Threadripper 1950X, 1920XAsus X399 ROG Zenith Extreme4x 8GB G.Skill Ripjaws V DDR4-3200 @ 2666 and 3200 MT/sIntel LGA 2066Intel Core i9-7900X, i7-7820XMSI X299 Gaming Pro Carbon AC4x 8GB G.Skill Ripjaws V DDR4-3200 @ 2666 and 3200 MT/sAMD Socket AM4 AMD Ryzen 7 1800XMSI X370 Xpower Gaming Titanium2x 8GB G.Skill Ripjaws V DDR4-3200 @ 3200 MT/sIntel LGA 1151 Intel Core i5-7700K MSI Z270 Gaming M72x 8GB G.Skill Ripjaws V DDR4-3200 @ 2666 and 3200 MT/sAll EVGA GeForce GTX 1080 FE 1TB Samsung PM863 SilverStone ST1500, 1500W Windows 10 Creators Update Version 1703Corsair H100i v2

MORE: Best CPUs

MORE: All CPUs Content

AotS: Escalation & Civilization VI

Ashes of the Singularity: Escalation

Ashes of the Singularity: Escalation responds well to the addition of more host processing resources, and that's good news for Core i9-7960X's 16 cores and high IPC throughput.

The -7960X provides great performance at stock settings and even better results after some tuning. We did, however, encounter lower minimum frame rates from -7960X compared to the Core i9-7900X, though. The same behavior manifested on our -7980XE, so we suspect this issue is endemic to Intel's HCC die. You can see where the frame rates dip in our performance over time chart. This is accompanied by higher frame times in those problem areas.

AMD's Ryzen Threadripper 1950X lags behind, though it does achieve a higher minimum frame rate and 99th percentile measurement. The Threadripper models also don’t suffer as much during the latter stages of our benchmark.

Civilization VI AI Test

The Civilization VI AI test measures performance in a turn-based strategy game. Intel's Core i9-7960X offers middling performance at its stock settings, but benefits greatly from our overclocking efforts. It even outpaces an overclocked Core i9-7900X.

Civilization VI Graphics Test

Core i9-7960X trails most of the field, including a stock Ryzen 7 1800X, in its out-of-box configuration. Overclocking propels it to the top of our chart, though.

The tuned Core i9-7960X does encounter some turbulence during the opening stages of our Civilization VI benchmark, though, manifesting as a spike in the frame time variance chart.

MORE: Best CPUs

MORE: All CPUs Content

Battlefield 1 & Dawn of War III

Battlefield 1

The Core i9-7960X fares better during our trudge across the O La Vittoria landscape and outpaces AMD's Threadripper and Intel's own -7900X at stock settings. Tuning provides a tangible performance boost and smoother gameplay, but overclocking also allows the Core i9-7900X to establish a razor-thin victory.

We do see quite a bit of variance from the Threadripper and Ryzen 7 models, particularly during the test's opening stage, which manifests as hitching on-screen.

Dawn of War III

The overclocked Core i9-7960X again captures the top spot after a bit of tuning. It also outperforms the Core i9-7900X at stock settings.

AMD's Ryzen Threadripper models again lag behind, though it's important to note that they're in the recommended Game Mode, halving thread count to minimize Infinity Fabric latency. This means the Ryzen Threadripper 1950X is doing battle with just 16 of its available 32 threads.

MORE: Best CPUs

MORE: All CPUs Content

Grand Theft Auto V, Hitman & Shadow of Mordor

Grand Theft Auto V

Grand Theft Auto V tends to perform best on Intel CPUs, and that pattern continues with our Core i9-7960X benchmarks, which outperforms the -7900X at stock and overclocked settings (though the delta between them is tiny at our highest overclocks).

Core i7-7700K fares well at its stock settings, but only slightly beats the stock Core i9-7960X (97.7 FPS) after an aggressive tuning session.

Hitman

Hitman scales well with additional processing resources, once again giving the -7960X a quantifiable advantage.

Shadow of Mordor

Shadow of Mordor is lightly threaded, which punishes the Core i9-7960X's low base frequency. We verified that this CPU sporadically boosts to 4.5 GHz during the test, but the speed-ups are so brief that they do little to improve our frame rate measurements. Even tuning doesn't help much. Notably, Intel's latest lags the overclocked -7900X by more than 5 FPS on average.

MORE: Best CPUs

MORE: All CPUs Content

Project CARS, Far Cry Primal & Rise of the Tomb Raider

Project CARS

The Core i9-7960X offers leading performance against the rest of our field at stock settings, and tuning yields a solid 11.5% gain on average.

Given its price, the quad-core Core i7-7700K continues to complement our GeForce GTX 1080 well, while AMD's Ryzen Threadripper processors suffer in Game mode. Then again, we know this title slows down even more if you leave Threadripper in Creator mode, so Intel's advantage is likely a simple matter of getting more done per cycle.

Far Cry Primal

Far Cry Primal refuses to load on CPUs with more than 24 threads. Although we could have disabled eight of the -7960X’s threads, we instead chose to turn off half of them to match Ryzen Threadripper 1950X’s thread count in Game mode as an experiment. This gives us a look at “like” performance with the same number of threads.

The Threadripper models don’t take a lead, but they aren’t far behind after tuning. Meanwhile, Intel's Core i7-7700K shows why we don't typically recommend high-end desktop CPUs for gaming PCs.

Rise of the Tomb Raider

The Core i9-7960X unsurprisingly delivers impressive performance during our Rise of the Tomb Raider benchmark, besting Core i9-7900X in stock and overclocked trim.

MORE: Best CPUs

MORE: All CPUs Content

Rendering, Encoding, Compression & Decompression

Web Browser

Dropping $1000+ for exceptional Web browsing performance is overkill, no doubt. But these tests do give us an idea of how responsive each CPU is using common desktop applications.

Our results line up based on clock rate and IPC throughput, it appears, highlighting the lightly threaded nature of many productivity-oriented workloads. Of course, those tasks do tend to favor Intel's architectures, which lead in every test.

Rendering

Heavy rendering workloads are easy fodder for high-end desktop CPUs, and Intel's Core i9-7960X delivers superior performance in almost every test we throw at it (except for the single-threaded ones).

The Kaby Lake-based Core i7-7700K outpaces Intel's Core i9-7960X in both single-threaded tests due to its higher Turbo Boost frequency and improved IPC throughput. Meanwhile, the tuned -7900X and -7820X flex their higher clock rates to carve out a lead in those same metrics.

Our stock -7960X beats the chip's overclocked configuration in the short single-threaded Cinebench test by virtue of its 4.4 GHz Turbo Boost 3.0 bin, which is slightly higher than the all-core 4.3 GHz we dialed in during our tuning session. In contrast, the single-threaded POV-ray benchmark caused Intel's Core i9-7960X to shed some frequency, likely due to increased thermal load during this longer test.

Encoding & Compression

A chart-topping core count proves beneficial to the Core i9-7960X during our HandBrake and compression/decompression workloads. But the $1700 CPU stumbles during the LAME benchmark. Intel has confirmed that its mesh architecture can lead to performance regressions in a few applications, and it appears that LAME might benefit from optimization for the new architecture.

PCMark 10

MORE: Best CPUs

MORE: All CPUs Content

Conclusion

Intel unabashedly aims its high-end desktop CPUs at those seeking the ultimate in performance, and from our first round of testing, it's apparent that Core i9-7960X delivers this across a range of heavily threaded workloads like rendering, encoding, and file compression. Aggressive Turbo Boost binning, which is far more ambitious than anything Broadwell-E offered, yields impressive results in lightly-threaded workloads too. Though there are rare exceptions, -7960X is also exceptional in games. It's no longer necessary to compromise alacrity in one discipline for devastating speed in another. The top Skylake-X CPUs simply crank up the clocks when their many cores aren't needed.

Of course, you have to pay a premium for this privilege. And if you aren't explicitly running well-threaded tasks on a regular basis, there's not much reason to spend a small fortune on Skylake-X. For gaming, we're using a geometric mean of the 99th percentile frame times, which we convert into an FPS measurement, to provide an easy-to-read performance outlook. The 99th percentile results are a good indicator of smoothness. Ryzen's extra cores could enable more performance in the future as software evolves to utilize them better, so we also include a chart with newer games that exploit host processing resources more thoroughly.

Enthusiasts who mostly plan on gaming: mainstream chips are still the smart choice.

It is remarkable that, for years, developers wrote code capable of utilizing a handful of cores, at most. Yet, in the span of a few months, we suddenly have processors with so many cores that some applications won’t even load. It’s easy to imagine that we’ll start seeing more software optimized for multi-core architectures in the years to come. AMD gets a lot of credit for this; the company's new processors obviously spurred Intel to respond with lower prices and faster processors.

But Intel still charges a big premium for access to its 16-core -7960X. You'll pay $700 more than Ryzen Threadripper 1950X or the 10-core Core i9-7900X. For many, an undeniable performance advantage might not be worth that extra money. But if every second you save translates to dollars, the higher price tag might be worth paying.

Although Intel isn't charging as much per core this generation, the company's segmentation practices persist. AMD gives you ECC memory support and more PCIe connectivity, for example. Meanwhile, Intel disables ECC support to dissuade its Xeon customers from adopting a less expensive platform, and it slowly shaves off PCIe lanes as you descend the Core hierarchy.

Grueling time constraints limited what we could do with Core i9-7960X on this first outing with the chip. But we do have more tests to run, and you can bet we'll revisit the exhaustive thermal/power analysis that went into our Core i9-7900X review. Plus, Core i9-7980XE is sitting here, waiting for a fixed motherboard firmware to truly unlock its potential; test results from Intel's flagship will follow shortly.

In the meantime, there's no question that Core i9-7960X is the fastest 16-core processor available. Intel is simply charging too much for the pleasure of owning one.

MORE: Best CPUs

MORE: All CPUs Content

AMD Ryzen Threadripper 1950X Game Mode, Benchmarked

palcorn@outlook.com (Paul Alcorn) — Mon, 18 Sep 2017 13:00:00 +0000

Finding The Right Modes

If you crave lots of cores and tons of PCIe connectivity, like most content creators, multitaskers, and software developers, then Threadripper is for you. It might also be a good fit if you're a gamer who simultaneously runs heavily threaded productivity applications in the background.

The Zeppelin die really is a feat of modern engineering. However, its architecture is dissimilar from anything that came before, creating issues in some software written prior to Ryzen's introduction. AMD worked with game developers to iron out the performance wrinkles we identified at launch, and we've seen big speed-ups in a number of titles as a result.

But expanding beyond Ryzen 7, 5, and 3 into a dual-die configuration adds a new set of challenges for Threadripper.

AMD's fix involves two toggles that affect how the processor operates, giving you modes optimized for whatever workload you're running. These switches create a total of four unique configurations to choose from. So, in a bid to condense the number of combinations, AMD created its Creator and Game modes.

Why Do We Need Game Mode?

The Zeppelin die consists of two quad-core CPU complexes (CCXes) woven together with the Infinity Fabric interconnect. Even in the single-die Ryzen 7/5/3 processors, this creates a layer of latency that affects communication between the CCXes. AMD builds upon that design with its Threadripper processors, leveraging two active Zeppelin dies. As you might imagine, this introduces another layer of Infinity Fabric latency.

Each die has its own memory and PCIe controllers. So, if a thread running on one core needs to access data resident in cache on the other die, it has to traverse the fabric between those dies and incur significant latency. Naturally, the latency penalty between dies is higher than it is between CCXes in the single-die configurations. To combat the potential for performance regression as a result of its "go-wide" approach, AMD devised an interesting solution: it introduced a new memory access switch that you can toggle via motherboard BIOS or the Ryzen Master software. The Local and Distributed settings flip between either NUMA (Non-Uniform Memory Access) or UMA (Uniform Memory Access).

NUMA works best if programs are designed specifically to utilize it. Even though most desktop PC software wasn't written with NUMA in mind, performance gains are still possible in non-NUMA applications.

Breaking Games

Some games simply won't load up when presented with Threadripper's 32 threads. That's right, AMD's flagship broke a few titles. The same thing will happen to Intel when its highest-end Skylake-X chips surface shortly.

Out of necessity, AMD created a Legacy Compatibility mode that executes a "bcdedit /set numproc XX" command in Windows. This command cuts the thread count in half. Fortunately, due to the operating system's default assignments, the command disables all of the cores/threads on the second die. That has a side benefit of eliminating thread-to-thread communication between disparate dies, solving the constant latency-inducing synchronization between threads during gaming workloads. It also prevents thread migration, lessening the chance of cache misses.

But What To Test?

The two new toggles give us a menu of options to mess with. AMD's Creator preset exposes 16C/32T and leaves the operating system in Distributed memory access mode. Those settings together should yield excellent performance in most productivity applications. Game mode cuts half the threads via compatibility mode and reduces memory and die-to-die latency with Local memory access.

As early as our first Ryzen review, we found that disabling SMT had a positive impact on some games. However, throwing in another setting to consider expands our list of viable configurations. Of course, some options make little sense for gaming. But they might be interesting to test with normal applications. We narrowed our list down based on Infinity Fabric measurements and prior experience with AMD processors.

Configurations	Local (NUMA) / Distributed (UMA)	Legacy Mode (On/Off)	SMT (Multi-Threading)
Creator Mode	Distributed	Off	On
Game Mode	Local	On	On
Custom - Local/SMT Off	Local	Off	Off
Custom - Local/SMT On	Local	Off	On

Disabling SMT in Game mode, which already halves core/thread count, would further cut into our available execution resources. So, we left SMT on and both dies active. We also had the option to leave all cores/threads active, but try isolating memory access to the local memory controller. We found that combination particularly useful in threaded games during our initial review. Of course, we also decided to test using the default Creator and Game modes.

Unfortunately, you can't just throw a switch and fire up your favorite game. Each change requires a reboot, chewing up precious time as you save open projects, halt conversations, and try to remember which web browser tabs to relaunch. Again, Game mode also halves thread count, which isn't good for running heavily threaded applications at full speed in the background while you play. Luckily, the breadth of available options should allow us to find a better configuration that leaves more processing resources at the ready.

AMD designed its Threadripper processors primary for use in enthusiast-oriented PCs, which tend to ship with high-resolution monitors. Higher resolutions are typically GPU-bound, reducing the variance between CPU modes. To better highlight performance trends and maximize our limited testing time, though, we chose to benchmark at 1920x1080, allowing us to spot the subtle differences between AMD's settings.

Overall, Game mode is designed to improve memory latency and help avoid excessive die-to-die Infinity Fabric traffic. Let's start with examining how our settings affect the Infinity Fabric and memory subsystem.

MORE: Best CPUs

MORE: All CPUs Content

AMD Ryzen Threadripper 1950XView Deal

Testing Ryzen's Infinity Fabric & Memory Subsystem

Infinity Fabric Latency And Bandwidth

The 256-bit Infinity Fabric crossbar ties the resources inside of a Zeppelin die together. Tacking on a second Zeppelin die to create Threadripper introduces another layer of the fabric, though. Cache accesses remain local to each CCX, but a large amount of memory, I/O, and thread-to-thread traffic still flow across that second layer.

It didn't take long for enthusiasts to figure out that AMD's Infinity Fabric is tied into the same frequency domain as the memory controller, so a memory overclock reduces latency and increases bandwidth through the crossbar. Performance in latency-sensitive applications (like games) consequently improves.

SiSoftware Sandra's Processor Multi-Core Efficiency test helps us illustrate the Infinity Fabric's performance. We use the Multi-Threaded metric with the "best pair match" setting (lowest latency). The utility measures ping times between threads to quantify fabric latency in every possible configuration.

The intra-core latency measurements represent communication between two logical threads resident on the same physical core, and as we can see, disabling SMT eliminates that measurement entirely. For the remaining setups, tuning reduces latency by a few nanoseconds. But this is attributable to higher clock rates. As we've seen in the past, increased memory frequencies have little effect on intra-core latency.

Intra-CCX measurements quantify latency between threads on the same CCX that are not resident on the same core. Increasing the clock rate yields larger ~6ns latency reductions.

Cross-CCX quantifies the latency between threads located on two separate CCXes, and we see a similar reduction thanks to overclocking. Notably, the Ryzen 7 1800X features much lower Cross-CCX latency than the stock Threadripper and most overclocked configurations. This is likely due to some form of provisioning, possibly in the scheduling algorithms, for Threadripper's extra layer of fabric.

As we can see, the overclocked Threadripper CPU in Game mode, which doesn't have an active fabric link to the other die, has the lowest Cross-CCX latency.

Die-To-Die measures communication between the two separate Zeppelin dies. Game mode effectively disables the second Zeppelin die at an operating system level, eliminating die-to-die latency entirely. The second die's uncore is still active though, which is necessary to ensure its I/O and memory controllers are still accessible.

Creator mode suffers the worst die-to-die latency, but tuning reduces it considerably. The two SMT options (on and off) receive large reductions from our overclocking efforts as well.

The utility measures fabric bandwidth too, which is critical for performance since data fetches from the remote memory also flow across the fabric. As such, AMD over-provisions the fabric and memory subsystem to optimize the distributed memory architecture.

Both the Creator mode and Local/SMT configurations offer the best fabric bandwidth, enjoying big boosts from overclocking. The Ryzen 7 1800X falls into the middle of the chart alongside Threadripper's Game mode, which is logical considering they are both effectively 8C/16T processors. Disabling SMT but leaving both dies active (Local/SMT off) yields a unique profile that provides higher performance with larger accesses and lower performance with smaller accesses.

Cache And Memory Latency

We tested with DDR4-2666 memory at stock settings and increased to DDR4-3200 for our overclocked configurations.

The Translation Look Aside Buffer is a cache that reduces access times by storing recently accessed memory addresses. Like all caches, the TLB has a limited capacity, so address requests that land in the TLB are "hits," while requests that land outside of the cache are "misses." Of course, hits are more desirable, and solid prefetcher performance yields higher hit rates.

Sequential access patterns are almost entirely prefetched into the TLB, so the sequential test is a good measure of prefetcher performance. The in-page random test measures random accesses within the same memory page. It also measures TLB performance and represents best-case random performance (this is the measurement vendors use for official spec sheets). The full random test features a mix of TLB hits and misses, with a strong likelihood of misses, so it quantifies worst-case latency.

Regardless of the memory access pattern, the smallest data chunks fit into the L1 cache. And as the size of the data increases, it populates the larger caches.

	L1	L2	L3	Main Memory
Range	2KB - 32KB	32KB - 512KB	512KB - 8MB	8MB - 1GB

Threadripper 1950X features better L2 and L3 latency than the Ryzen 7 1800X with every type of access pattern. Also, we spot notable latency reductions via overclocking for Threadripper's L1, L2, and L3 caches.

That changes as the workload flows out to main memory. Threadripper's Creator mode (the default setting) has the highest latency with every access pattern. This is a direct result of memory accesses landing in the remote memory. Our in-page measurements mirror AMD's 86.9ns specification, but worst-case full random access exceeds 120ns. Overclocking the processor and memory lowers latency, but Creator mode still doesn't overtake any of the configurations we compare it to.

Switching into NUMA mode with the Local setting improves main memory access dramatically for the other configurations. We measure ~60ns for in-page near memory access, again in line with AMD's specifications, while worst-case latency weighs in at 100ns.

Cache Bandwidth

Each CCX has its own caches, so a Threadripper CPU features four distinct clusters of L1, L2, and L3 memory. Our bandwidth benchmark illustrates the aggregate performance of these tiers.

During the single-threaded test, Ryzen 7 1800X demonstrates lower throughput than the Threadripper processors. The other configurations clump together in familiar stock and overclocked groups.

The multi-threaded tests are more interesting; we see Ryzen 7 1800X and the two Threadripper Game modes fall to the bottom of the chart. Because Game mode disables the cores on one die, it effectively takes the corresponding cache out of commission.

MORE: Best CPUs

MORE: All CPUs Content

AMD Ryzen Threadripper 1950XView Deal

VRMark & 3DMark

Test Notes

We used the same test platforms and settings outlined in our recent AMD Ryzen Threadripper 1920X Review. AMD's gaming performance is a moving target that continues improving over time. So, today's story reflects all processors re-tested with the latest chipset, BIOS, GPU drivers, and game patches.

Configurations	Local (NUMA) / Distributed (UMA)	Legacy Mode (On/Off)	SMT (Multi-Threading)
Creator Mode	Distributed	Off	On
Game Mode	Local	On	On
Custom - Local/SMT Off	Local	Off	Off
Custom - Local/SMT (on)	Local	Off	On

VRMark & 3DMark

We aren't big fans of using synthetic benchmarks to measure game performance, but 3DMark's DX11 and DX12 CPU tests provide useful insight into the amount of horsepower available to game engines.

Futuremark's VRMark test lets you gauge your system's suitability for use with the HTC Vive or Oculus Rift, even if you don't currently own an HMD. The Orange Room test is based on the suggested system requirements for current-generation HTC Vive and Oculus Rift HMDs. Futuremark defines a passing score as anything above 109 FPS.

VRMark responds well to high IPC throughput and frequency. Overclocking in Game mode yields the best performance, offering better results than overclocking under Creator mode.

The DX12 tests favors higher thread counts, so Creator mode becomes desirable.

Futuremark's threaded DX11 physics test responds well to the full complement of cores and threads, so the overclocked Local/SMT configuration dominates. Data locality, enforced through the Local setting, is the only difference between Creator mode and Local/SMT, so it appears the DX11 test favors this.

The API test also lends itself to high core counts. Both the Creator mode and Local/SMT configuration provides access to Threadripper 1950X's 32 threads. However, the Local/SMT configuration suffers during the DX12 and Vulkan tests.

MORE: Best CPUs

MORE: All CPUs Content

AMD Ryzen Threadripper 1950XView Deal

AotS: Escalation & Civilization VI

Ashes of the Singularity: Escalation

Running overclocked in Creator mode narrowly yields the best average frame rate during the Ashes of the Singularity: Escalation benchmark

Game mode reduces the core/thread count and available cache, which unsurprisingly results in low performance during this heavily threaded title. The benchmark does appear to favor physical cores over logical processors, as the Local/SMT Off configuration beats the Local/SMT option with all of the architecture's threads accessible.

Civilization VI AI Test

The stock Core i9-7900X is faster than AMD's Threadripper 1950X, regardless of how we configure it. Tuning up to 3.9 GHz at least makes the Threadripper chip competitive.

This benchmark tends to favor physical cores, but we see some jockeying between the stock and overclocked SMT on/off configurations. Notably, Game mode provides impressive performance once we overclock. But it falls to the bottom of our chart in the CPU's stock form.

Civilization VI Graphics Test

The Civilization VI graphics test finds the Local/SMT configuration leading narrowly at stock and overclocked settings. Considering the gulf between Creator mode and the NUMA-enabled configurations (Game mode, Local/SMT On, and Local/SMT), this test appears to favor localized memory access.

Intel's Core i9-7900X suffers due to some of the performance regressions associated with Skylake-X's mesh architecture.

MORE: Best CPUs

MORE: All CPUs Content

AMD Ryzen Threadripper 1950XView Deal

Battlefield 1, Dawn of War III & Grand Theft Auto V

Battlefield 1 (DX11)

Battlefield 1 benefits from disabling SMT, which enables the best overall performance in both stock and overclocked configurations. The title does not respond well to AMD's Game mode, which shows up in the chart's lowest position. Tuning only helps a little bit. Creator mode performs surprisingly well in both stock and overclocked configurations.

Warhammer 40,000: Dawn of War III

Game mode yields the best performance at stock and overclocked settings.

The Local/SMT options (on and off) both offer similar performance, with Local/SMT taking a slight lead. Notably, Creator mode suffers from reduced performance in both configurations; even tuning can't give it an edge over Game mode at stock clock rates.

Grand Theft Auto V

Grand Theft Auto V responds well to Game mode in both stock and overclocked configurations, while the Local/SMT configurations offer similar average frame rates. Creator mode trails the other configurations again, but the performance delta isn't as pronounced as we've seen in other titles.

MORE: Best CPUs

MORE: All CPUs Content

AMD Ryzen Threadripper 1950XView Deal

Hitman, Shadow of Mordor & Project CARS

Hitman (2016)

The Local/SMT on/off configurations offer similar performance at stock clock rates. But the delta between them widens in favor of Local/SMT Off after we do a bit of overclocking.

Intel's Core i9-7900X offers the best performance, but it also causes the most frame time variance during our benchmark. This seems attributable to blips as scenes change in the test sequence. To its credit, the -7900X provides the best 99th percentile frame rate results.

Middle-earth: Shadow of Mordor

Shadow of Mordor becomes more graphics-bound as we drop in higher-end CPUs, so there is little variation between most of today's contenders. The exception is Threadripper in Creator mode, which struggles with older, lightly-threaded titles. Game mode predictably fares well during the test, though it's roughly equivalent to our Local/SMT configuration.

Project CARS

The Madness game engine was designed to facilitate parallelization, but it doesn't respond well to Creator mode. This performance disparity is likely due to the distributed memory access; most games just aren't accustomed to this type of architecture.

There goes our theory that Creator mode offers the best performance in well-threaded games. Once again, the Local/SMT off combination lands in the middle of our chart, while Local/SMT switches on all 32 threads to secure a victory.

MORE: Best CPUs

MORE: All CPUs Content

AMD Ryzen Threadripper 1950XView Deal

Far Cry Primal, Rise of the Tomb Raider & The Witcher 3: Wild Hunt

Far Cry Primal

Far Cry Primal will not load when it detects more than 20 threads, so you have to flip the "Legacy" toggle and reboot to reduce the number of threads active on Threadripper. Disabling SMT also reduces the thread count, so you can play the game using our Local/SMT Off configuration.

This game's Dunia Engine 2 tends to perform better with SMT disabled anyway, so it's surprising to see AMD's Game mode take the lead.

Interestingly, Core i9-7900X offers the best average frame rate, but suffers some frame time disturbance at the beginning and end of our benchmark, while Threadripper provides a smooth experience in both configurations. The overclocked Threadripper Game mode configuration also offers the best 99th percentile measurement.

Rise of the Tomb Raider

Rise of the Tomb Raider was once a less-than-ideal showcase of Ryzen's capabilities. But a recent patch fixed most of the CPU family's issues.

Creator mode trails at stock settings, but takes the lead among AMD's Threadripper processors after some overclocking. An overclocked -7900X can push to 150 FPS in this title, so a graphics bottleneck isn't the easy answer. There is almost no variation between the leading Threadripper configurations, indicating that some other limitation is affecting our results. In either case, the overclocked Threadripper's 99th percentile measurements are within range of the -7900X, and the processors do not suffer any significant hitching or stuttering.

The Witcher 3: Wild Hunt

A graphics bottleneck looms during the Witcher 3: Wild Hunt benchmark, but it also allows us to examine performance during a heavy scene transition at the beginning of our jog through the woods. Both Game and Creator modes stutter noticeably during this sequence, while the Local/SMT configuration doesn't stutter as badly. The Local/SMT Off setting also experiences some variance toward the test's end.

MORE: Best CPUs

MORE: All CPUs Content

AMD Ryzen Threadripper 1950XView Deal

Final Analysis

AMD created new settings to help customers optimize the performance of its unique architecture. The options enable capabilities that simply weren't available on the desktop before now. Naturally, it's important to flip those switches and see what happens, particularly given the broad range of workloads that AMD says its Threadripper CPUs excel in.

AMD aims Threadripper at content creators, heavy multitaskers, and gamers who stream simultaneously. It also says the processors are ideal for gaming at high resolutions. Ryzen Threadripper 1950X isn't intended for playing around at low resolutions, particularly in older, lightly-threaded titles. Still, we tested at 1920x1080 to emphasize the difference between the various modes and settings AMD exposes, rather than show you results bound by our graphics card.

We plotted performance with both average frame rates and a geometric mean of the 99^th percentile frame times (a good indicator of smoothness), which we convert into an FPS measurement. Our suite includes six games released in 2016 and five older titles that launched in 2014/2015. Threadripper’s extra cores could enable more performance in the future as software evolves to utilize them better, so we also include a chart with newer games that exploit host processing resources more thoroughly.

Looking at the results spread across our entire gaming suite, the difference in performance between modes is surprisingly small. Four frames per second, on average, separate the fastest and slowest settings at stock clock rates. And we see a mere ~2 FPS delta with 99th percentile frame times. Overclocking the processor offers a more significant jump in performance than any mode configuration, and the difference between the various overclocked configurations is also small. We do see slightly larger deltas between the configurations in new titles, but again, they certainly aren't huge. And of course, gaming at higher resolutions does even more to narrow the gaps between settings.

Creator mode exhibits the bipolar gaming performance we've come to expect; it trails by a significant margin in lightly threaded games, but excels in titles that utilize lots of threads and uniform memory access. Unfortunately, some games appear to suffer from the Distributed mode, so leading performance in threaded titles isn't a given. This mode stumbles hard enough during our Civilization VI, Warhammer, Grand Theft Auto V, and Project CARS benchmarks that you'll certainly want to reboot and switch out of it.

Game mode provides solid performance in lightly threaded titles and lower frame rates in more heavily threaded games, but it often gives you a better 99th percentile result, indicating a smoother experience. It is a more reliable default mode, though you'll want to evaluate the experience for yourself on a case-by-case basis.

Switching off SMT and leaving both dies active, as we did with our Local/SMT Off configuration, provides good results, even as it cuts into available host processing resources.

Using Game mode or disabling SMT doesn't make much sense to us. A big part of Threadripper's allure is its ability to hammer through threaded workloads, even as other tasks are running simultaneously. Switching into Game mode sacrifices threads and cache, giving up performance in those situations.

Beyond recommending against Creator mode for gaming, we can't simply suggest one "best" combination of settings for single- and multi-threaded titles. There is simply too much variance as we expand the breadth of our suite. And that's likely why AMD enabled as many options as it did. Unfortunately, characterizing the behavior of your favorite games and rebooting between sessions probably won't be a popular pastime among enthusiasts.

Compromise might be the best approach. Leaving all threads active and switching to local memory access seems to provide the best of both worlds. This mode attempts to pin memory to the die executing the workload, thus offering a decent performance boost. But it also leaves the 1950X's complement of 32 threads available for heavy processing. Overall, this gives you optimal performance.

Of course, there are a few titles that won't initialize when confronted with 32 threads, so Legacy mode is a requirement in those isolated cases. It will be interesting to see how Intel tackles the same issue when its high-end Skylake-X models land.

Even at its worst, Threadripper delivers adequate performance in a majority of games at 1080p. Most enthusiasts will pair these high-end CPUs with fast graphics cards and high-res monitors though. In that case, Threadripper is easily quick enough to keep up at graphics-bound settings. Having 32 threads at your disposal for heavy lifting in the background is nice, too.

MORE: Best CPUs

MORE: All CPUs Content

AMD Ryzen Threadripper 1950XView Deal

Retail AMD Threadripper Delidded, Four Die Present

palcorn@outlook.com (Paul Alcorn) — Sat, 16 Sep 2017 20:00:00 +0000

Der8auer purchased a retail Threadripper 1950X and found what appears to be four active Zeppelin die soldered to the IHS. We followed up with AMD for more detail.

Earlier this year, Der8auer delidded a Threadripper engineering sample and found that the processors still feature all four Zeppelin die, similar to AMD's EPYC data center processors. AMD uses only two die for the Threadripper models, so many speculated that it might be possible to unlock the other dies and their extra cores.

AMD responded that two of the die on shipping Threadripper models are not the same as those Der8auer found in the engineering sample--the additional two die are "dummy" die that the company uses to provide structural stability for the package. These additional die prevent the IHS (Integrated Heat Spreader) from caving in when you tighten down the heatsink.

Recently, Der8auer followed up and purchased a retail Threadripper 1950X, then created a custom delidder to remove the IHS. Again, he found four Zeppelin die soldered to the IHS but decided to take his experiment further and determine if the additional packages were "real" die or just inserts. Der8auer removed the die by placing the processor onto a heated copper block (300C), thus weakening the bonds from the PCB. After removing the packages, he used a diamond lapping material to expose the die.

Der8auer found that all four die appear to be "real" die (you can clearly see the two four-core complexes). We followed up with AMD and inquired if the die were functional. AMD responded that two of the die are in fact non-functional "dummy" die that aren't connected electrically to the substrate.

During AMD's Hotchips presentation, the company outlined Threadripper's MCM package and indicated that each die features four Infinity Fabric links, but the company uses only three links for the EPYC processors. This technique minimizes trace length to reduce latency. Because the company doesn't electrically connect two of the Threadripper die to the substrate, it's natural to assume that the company is using only two Infinity Fabric links for the two active dies. That also scuttles any chances of activating the two "extra" die.

Mounting dies to a substrate is like any other manufacturing process; there can be defects. It's possible that the Threadripper processors are simply EPYC processors with die that weren't successfully mated to the substrate. Conversely, it's also possible that AMD is using defective die for the fillers, but in either case, it's doubtful that the company is wasting functioning silicon. We've followed up with AMD for additional clarification and will update as necessary. In the meantime, we suggest watching the video for an interesting look inside the die removal process.

Ryzen Threadripper 1900X Available, Bootable NVMe RAID Support Comes To X399

palcorn@outlook.com (Paul Alcorn) — Thu, 31 Aug 2017 13:25:00 +0000

Hot on the heels of AMD's Ryzen Threadripper 1950X and 1920X launch, the company shared more information about its Threadripper 1900X model that is available for purchase today. AMD gears the new entrant for prosumers, creators, developers, and researchers, and of course, multi-tasking gamers. The company also announced support for bootable NVMe RAID arrays on X399 motherboards.

AMD unabashedly aims the Threadripper lineup at the HEDT market with copious core counts for demanding workloads. The series drops into premium X399 motherboards with the TR4 socket, and the 1900X slots in as the 8C/16T model. AMD also has eight-core Ryzen 7 models paired with mainstream AM4 motherboards. Aside from the similar core counts, the 1800X is much different than its larger 1900X cousin, largely due to architectural and platform differences.

The Ryzen 7 series is a single-die design, whereas the Threadripper 1900X features a dual-die design with four active cores in each die, along with two dummy die to ensure package stability. AMD selects the top 5% of its die for Threadripper models, so the 1900X also features higher base and turbo frequencies than the 1800X, but the most important differences lay in connectivity options.

The 1900X features two dual-channel memory controllers that provide access to quad-channel memory (for up to 512GB of memory capacity), a significant gain over the 1800X's dual-channel memory. AMD's less stringent segmentation results in the full complement of connectivity options, even for its lower core count models. The 1900X supports up to 64 PCIe lanes (the 1800X's 16 lanes pale in comparison) and more USB ports. The improved memory performance provides a big boost for memory-bound workloads, and increased PCIe connectivity makes for a compelling offering for users who have more elaborate requirements.

	Threadripper 1950X	Core i9-7900X	Threadripper 1920X	Core i7-7820X	Threadripper 1900X	Ryzen 7 1800X
Price	$999	$999	$799	$599	$549	$499
Interface/Chiset	TR4 / X399	LGA2066 / X299	TR4 / X399	LGA2066 / X299	TR4 / X399	AM4 / X370
Cores/Threads	16/32	10/20	12/24	8/16	8/16	8/16
TDP	180W	140W	180W	140W	180W	95W
Base Frequency (GHz)	3.4	3.3	3.5	3.6	3.8	3.6
Boost Frequency (GHz)	4.0 (4.2 XFR)	4.3 / 4.5 (TB 3.0)	4.0 (4.2 XFR)	4.3 / 4.5 (TB 3.0)	4.0 (4.2 XFR)	4.0 (4.1 XFR)
L3 cache (L2+L3)	40MB	23.75MB	38MB	19MB	20MB	20MB
Memory Support	DDR4-2667	DDR4-2666	DDR4-2667	DDR4-2666	DDR4-2667	DDR4-2667
Memory Controller	Quad Channel	Quad Channel	Quad Channel	Quad Channel	Quad Channel	Dual Channel
Unlocked Multiplier	Yes	Yes	Yes	Yes	Yes	Yes
PCIe Lanes	64	44	64	28	64	16

The Ryzen 7 series comes with eight cores and varying clock frequencies spread across three models that range from $329 to $499. Due to AMD's unlocked multipliers, the three Ryzen 7 models provide nearly identical performance if you overclock. We often see the Ryzen 7 models selling far below MSRP, so they certainly represent a great value. The Ryzen 7 and 5 series will continue to serve gamers well, including those who stream, whereas the 1900X slots in for users who need access to more sophisticated use cases. It also provides a lower cost of entry for the TR4 platform, which provides a cheaper upgrade path in the future.

AMD provided several performance comparisons (test notes at the end of this article) that compare the 1900X to its more well-equipped siblings. As expected, the 1900X features better performance than the 1950X and 1920X in lightly threaded titles, while it loses a minor amount of performance in games that favor more cores. The 1900X also fares well compared to the 1800X in games that benefit from higher clock speeds, whereas the 1800X holds the advantage in threaded titles. AMD's benchmarks compare the 1920X in legacy mode, which halves the number of threads, essentially creating a four-core processor. This mode sidesteps the latency of inter-die communication, but we can see that it has an impact on performance in some games.

The story is much different when we shift focus to productivity applications. The 1900X's superior memory subsystem facilitates more performance across the board in applications like POVRay, Blender, Handbrake, and Veracrypt.

Overall, the 1900X's benefits boil down to its better connectivity options. AMD outlined several use cases that would benefit from that, such as using a GPU, streaming card, and NVMe RAID array simultaneously. This setup consumes only 48 of Threadripper's 64 lanes, so AMD also outlined a few more stringent setups for deep learning and render farms. For instance, the 1900X accommodates up to four GPUs (although two will have restricted bandwidth) and an NVMe array.

Speaking of NVMe, the X399 platform launched without support for bootable NVMe RAID, but the company is adding support for RAID 0, 1, and 10 via firmware updates. You can build bootable RAID arrays of up to ten devices, and the upgrade is free.

Thoughts

Aside from the step down to eight cores, the 1900X is quite similar to its Threadripper predecessors. Pairing a less expensive chip with an expensive motherboard can be a bad decision, much like we see with Intel's Core i7 and i5 processors on the X299 chipset. Those processors offer roughly the same performance as their mainstream counterparts, and because you don't gain any extra connectivity options to make it worthwhile, shifting to the expensive X299 platform offers zero benefit to users.

AMD's approach is entirely different. It's eight-core Ryzen Threadripper 1900X features a different architecture, and pairing it with X399's expanded connectivity options yields a tangible benefit over the mainstream Ryzen 7 series. The 1900X surely isn't for everyone, but it should be a good fit for users with especially specific requirements.

The Ryzen Threadripper 1900X is available for order today, and NVMe RAID support comes September 25.

AMD Ryzen Threadripper 1900XView Deal

AMD Ryzen Threadripper 1920X Review

palcorn@outlook.com (Paul Alcorn) — Wed, 30 Aug 2017 13:00:00 +0000

Introduction

AMD’s Ryzen Threadripper 1950X put a feather in the cap of its high-end desktop aspirations. The company's new line-up challenges Intel's best efforts. But, as usual, some of the best value in AMD's product stack is found in the mid-range models. Ryzen Threadripper 1920X comes arms with 12 physical cores and SMT, enabling 24 concurrent threads fed by 38MB of cache, a quad-channel memory controller, and 64 lanes of PCIe. All of that costs $800, dramatically undercutting the 10-core Core i9-7900X.

Based on the back-and-forth we've witnessed this year, it appears the Ryzen family of CPUs may have caught Intel off-balance. AMD's siren call to enthusiasts includes lower prices, more cores, less segmentation, soldered heat spreaders, less expensive motherboards, and a longer commitment to each platform.

Intel does have pricier Skylake-X options available, but they sag under the weight of deliberate segmentation that fuses off native features on the cheaper models. Don't count Intel out, though; its beefiest Skylake-X chips are still forthcoming, along with a salvo of mainstream Coffee Lake CPUs to rival Ryzen 7, 5, and 3.

AMD has an aggressive roadmap it'll use to improve the Zen architecture and transition to smaller nodes, so the company should remain a competitive force to be reckoned with. Ryzen Threadripper 1920X is a great start, though. Based on the 1950X we already reviewed, this processor is expected to perform well at a reasonable price point (plus the highest overclocking ceiling we’ve seen on a Ryzen processor).

Meet Ryzen Threadripper 1920X

AMD designed its Threadripper processors for anyone able to utilize lots of cores and tons of PCIe connectivity. Think content creators, heavy multi-taskers, and software developers.

The 12C/24T Threadripper 1920X features a 3.5 GHz base clock, which is just 100 MHz higher than the 16C/32T 1950X. Surprisingly, the two chips share the same 3.7 GHz boost frequency for heavily-threaded workloads and a four-core 4 GHz setting for less taxing tasks. If your cooler is robust enough, both processors also enable a four-core 4.2 GHz XFR ceiling.

Like all of AMD's Ryzen processors, the 1920X utilizes two quad-core complexes combined into a single Zeppelin die. Two Zeppelin dies, tied together using the Infinity Fabric interconnect into a multi-chip module, come together to create Threadripper CPUs wielding 16 physical cores. AMD creates the 12-core 1920X by disabling four of them, leaving six cores per die (3+3).

The disabled cores serve as dark silicon, which absorbs heat dissipated by the active circuitry. This, coupled with AMD's use of a soldered heat spreader and aggressive binning (the company claims to use the top 5% of Zeppelin dies), leads to impressive overclocking headroom from our 1920X sample. We maintained a 4.1 GHz overclock, the highest achieved with any Ryzen CPU in our U.S. lab, using a relatively tame 1.42V.

	Threadripper 1950X	Core i9-7900X	Threadripper 1920X	Core i7-7820X	Threadripper 1900X
Price	$1000	$1000	$800	$600	$550
Interface/Chipset	TR4 / X399	LGA2066 / X299	TR4 / X399	LGA2066 / X299	TR4 / X399
Cores/Threads	16/32	10/20	12/24	8/16	8/16
TDP	180W	140W	180W	140W	180W
Base Frequency (GHz)	3.4	3.3	3.5	3.6	3.8
Boost Frequency (GHz)	4.0 (4.2 XFR)	4.3 / 4.5 (TB 3.0)	4.0 (4.2 XFR)	4.3 / 4.5 (TB 3.0)	4.0 (4.2 XFR)
Cache (L2+L3)	40MB	23.75MB	38MB	19MB	20MB
Memory Support	DDR4-2667	DDR4-2666	DDR4-2667	DDR4-2666	DDR4-2667
Memory Controller	Quad-Channel	Quad-Channel	Quad-Channel	Quad-Channel	Quad-Channel
Unlocked Multiplier	Yes	Yes	Yes	Yes	Yes
PCIe Lanes	64	44	64	28	64

Ryzen Threadripper 1920X slots into the large price gap between Intel's Core i9-7900X and $600 i7-7820X. Making AMD's solution more interesting is the fact that Intel cuts PCIe connectivity from 44 lanes to 28 as you drop to the Core i7. In comparison, the Threadripper chip boasts 64 lanes, though four are reserved for AMD's chipset. The extra I/O comes in handy for multi-GPU configurations, large PCIe-based storage arrays, and streamers using dedicated capture cards.

The 1920X and 1950X both feature 32MB of L3 cache sliced into 16MB per Zeppelin die. You do lose 2MB of L2 cache to the four disabled cores, leaving 512KB per core, or 6MB across the MCM, active. Despite the disabled cores and cache, AMD still rates its 1920X with a 180W TDP.

Enthusiasts have to love that AMD uses Indium solder instead of the thermal paste Intel employs. Threadripper's large IHS helps with heat too, and the chip generally features solid thermal performance. We haven't encountered any serious heat concerns with the Threadripper models, which we can't say for Intel's Skylake-X CPUs.

Ryzen Threadripper Memory Support	MT/s
Quad-Channel/Dual-Rank/Two DIMMS per Channel (8)	1866
Quad-Channel/Single-Rank/Two DIMMs Per Channel (8)	2133
Quad-Channel/Dual-Rank/One DIMM Per Channel (4)	2400
Quad-Channel/Single-Rank/One DIMM Per Channel (4)	2677

Threadripper features independent dual-channel memory controllers, one paired with each die, that combine to provide quad-channel support with varying data transfer rates based upon memory types and DIMMs per channel. The platform supports ECC memory and a functional limit of 256GB of DDR4, though it can support up to 2TB as memory capacity increases.

The distributed memory alignment, along with the latency imposed by traversing the fabric between two separate dies, creates unique challenges for applications sensitive to timing. AMD has also discovered that certain games won't fire up with all of Threadripper's cores enabled. So the company implemented a pair of toggles that enable either UMA or NUMA mode to mitigate memory latency concerns, and a Legacy setting to disable one CCX, solving compatibility issues.

Selectable Creator and Game modes provide users with performance profiles tailored for either content creation or gaming. We covered how the underlying architecture responds to these modes in our AMD Ryzen Threadripper 1950X review.

We are starting to see dedicated coolers trickle out from leading vendors for AMD's massive 4094-pin TR4 socket. In the interim, AMD also includes an Asetek bracket with all Threadripper models to provide widespread compatibility with existing closed-loop coolers from several vendors.

MORE: Best CPUs

MORE: Intel & AMD Processor Benchmarks Hierarchy

MORE: All CPUs Content

How We Test

Comparison Products

Ryzen 7 1800X

Intel Core i9-7900X

Intel Core i7-7820X Skylake-X

Test Systems

We introduced our new test system and methodology in How We Test Graphics Cards. If you'd like more detail about our general approach, check that piece out.

Test System & Configuration
Hardware	Germany AMD Socket SP3 (TR4)AMD Ryzen Threadripper 1950X, 1920XAsus X399 ROG Zenith Extreme4x 8GB G.Skill RipJaws V DDR4-3200Intel LGA 2066Intel Core i9-7900XMSI X299 Gaming Pro Carbon AC4x 4GB G.Skill RipJaws IV DDR4-2600AMD Socket AM4 WorkstationAMD Ryzen 7 1800X, 1700X, 1600X, 1500XMSI X370 Tomahawk4x 8GB G.Skill TridentZ DDR4-3200Intel LGA 2011v3Intel Core i7-6900KMSI X99S XPower Gaming Titanium4x 4GB Crucial Ballistix DDR4-2400Intel LGA 1151Intel Core i7-7700KMSI Z270 Gaming 72x 8GB Corsair Vengeance DDR4-3200 @ 2400 MT/sAll SystemsGeForce GTX 1080 Founders EditionNvidia Quadro P6000 (Workstation)1x 1TB Toshiba OCZ RD400 (M.2, System)2x 960GB Toshiba OCZ TR150 (Storage, Images)be quiet! Dark Power Pro 11, 850W Power SupplyWindows 10 Pro (Creators Update)U.S.AMD Socket SP3 (TR4)AMD Ryzen Threadripper 1950X, 1920XAsus X399 ROG Zenith Extreme4x 8GB G.Skill RipJaws V DDR4-3200 @ 2666 and 3200 MT/sIntel LGA 2066Intel Core i9-7900X, i7-7820XMSI X299 Gaming Pro Carbon AC4x 8GB G.Skill RipJaws V DDR4-3200 @ 2666 and 3200 MT/sAMD Socket AM4 AMD Ryzen 7 1800XMSI X370 Xpower Gaming Titanium2x 8GB G.Skill RipJaws V DDR4-3200 @ 3200 MT/sIntel LGA 1151 Intel Core i5-7700K MSI Z270 Gaming M72x 8GB G.Skill RipJaws V DDR4-3200 @ 2666 and 3200 MT/sAll EVGA GeForce GTX 1080 FE 1TB Samsung PM863 SilverStone ST1500, 1500W Windows 10 Creators Update Version 1703
Cooling	GermanyAlphacool Eiszeit 2000 ChillerAlphacool Eisblock XPXThermal Grizzly Kryonaut (For Cooler Switch)
Monitor	Eizo EV3237-BK
PC Case	Lian Li PC-T70 with Extension Kit and Mods Configurations: Open Benchtable, Closed Case
Power Consumption Measurement	Contact-free DC Measurement at PCIe Slot (Using a Riser Card) Contact-free DC Measurement at External Auxiliary Power Supply Cable Direct Voltage Measurement at Power Supply 2x Rohde & Schwarz HMO 3054, 500MHz Digital Multi-Channel Oscilloscope with Storage Function4x Rohde & Schwarz HZO50 Current Probe (1mA - 30A, 100kHz, DC) 4x Rohde & Schwarz HZ355 (10:1 Probes, 500MHz) 1x Rohde & Schwarz HMC 8012 Digital Multimeter with Storage Function
Thermal Measurement	1x Optris PI640 80Hz Infrared Camera + PI Connect Real-Time Infrared Monitoring and Recording
Acoustic Measurement	NTI Audio M2211 (with Calibration File, Low Cut at 50Hz) Steinberg UR12 (with Phantom Power for Microphones)Creative X7, Smaart v.7 Custom-Made Proprietary Measurement Chamber, 3.5 x 1.8 x 2.2m (L x D x H) Perpendicular to Center of Noise Source(s), Measurement Distance of 50cm Noise Level in dB(A) (Slow), Real-time Frequency Analyzer (RTA) Graphical Frequency Spectrum of Noise

MORE: Best CPUs

MORE: All CPUs Content

VRMark, 3DMark & AotS: Escalation

Test Notes

Intel released new microcode for its Skylake-X processors recently, which reduces performance in some titles and lowers the AVX offset by two bins. We also noticed far lower Turbo Boost activation thresholds, though that could separately be the result of MSI's newest BIOS. The changes likely come in response to some of the power and thermal issues we encountered during our extended testing. We consequently retested both Skylake-X processors with the newest microcode.

AMD's Ryzen gaming performance is also a moving target, though it continues to improve over time. Today's story reflects all processors re-tested with the latest chipset, BIOS, GPU drivers, and game patches. We continue seeking out the best performance possible, so in today's review, we dial in AMD's Game mode for our game benchmarks.

VRMark & 3DMark

We aren't big fans of using synthetic benchmarks to measure game performance, but 3DMark's DX11 and DX12 CPU tests provide useful insight into the amount of horsepower available to game engines.

VRMark responds well to high IPC throughput and frequency, so it's not surprising to see the 1920X benefit from a clock rate advantage over the 1950X at stock settings. Extra overclocking headroom gives the 1920X a bigger boost once we start tuning.

The Intel processors lead, but it's possible that Ryzen-specific optimizations could improve Threadripper's results.

We also tested the 1920X and 1950X in Creator mode for the threaded 3DMark DX11 and DX12 tests. Big core counts propel Threadripper to the top of our charts when it's overclocked, but Game mode halves the number of available threads, causing lackluster performance. Notably, both overclocked Threadripper models in Creator mode yield the best performance.

The Vulkan API responds exceedingly well to Threadripper's architecture, and the tuned 1920X delivers excellent DX11 single-threaded performance, particularly in game mode.

Ashes of the Singularity: Escalation

Switching the 1920X into Game mode creates a 6C/12T configuration that maximizes memory locality and eliminates die-to-die latency. But that comes at the expense of performance in parallel workloads. As such, Threadripper falls below its 8C/16T Ryzen 7 1800X counterpart in this heavily-threaded game.

We include an additional slide with test results using various settings. These numbers highlight that Game mode has a positive impact on lightly-threaded and otherwise incompatible titles, but can be a hindrance for more taxing workloads. A bit of tuning (and switching to Creator mode) pushes the Threadripper models above Intel's Core i9-7900X.

MORE: Best CPUs

MORE: All CPUs Content

Civilization VI, Battlefield 1 & Dawn of War III

Civilization VI AI Test

High frequencies and IPC throughput provide the best results in Civilization VI's AI test, which measures the available computational horsepower during a turn-based strategy gaming session.

The overclocked Intel processors lead, while AMD's Ryzen models populate the lower half of our chart. Interestingly, though, Threadripper 1920X triumphs over the 1950X in yet another lightly threaded title.

Civilization VI Graphics Test

Both Skylake-X-based processors fall to the bottom of our chart in their stock configuration, which doesn't make sense given superior host processing resources. This is one of the notable performance regressions we have encountered with Intel’s new mesh architecture.

The Threadripper models lead the Intel competition at stock and overclocked settings. But the nimble Ryzen 7 1800X also demonstrates the value of a much less expensive platform. Core i7-7700K also does well out of the box, managing to outpace the brawnier stock Threadripper models. However, both the -7900X and -7700K do suffer significant frame time outliers.

Battlefield 1 (DX11)

The less expensive Core i7-7820X and -7700K CPUs outpace AMD's Threadripper processors, though much of this is due to Ryzen's tendency to stutter during the opening seconds of our benchmark sequence.

The 1950X outperforms Threadripper 1920X, suggesting this title scales with additional cores.

Warhammer 40,000: Dawn of War III

A stock Core i7-7700K dominates this test. Still, the overclocked Threadripper models put up a good fight as they deliver a smooth experience at over 100 FPS.

The 1920X leads over the 1950X in both stock and overclocked configurations. Meanwhile, Intel's Core i7-7820X suffers severe frame time variance that manifests as visible stuttering.

MORE: Best CPUs

MORE: All CPUs Content

Grand Theft Auto V, Hitman & Shadow of Mordor

Grand Theft Auto V

We measure performance during Grand Theft Auto V's F-16 flight sequence with the built-in benchmark.

Grand Theft Auto V has always responded well to Intel's architectures. But a tuned 1950X in Game mode fares well, rivaling the Core i9-7900X and outpacing the -7700K.

The tuned 1920X isn't far behind the -7700K, Meanwhile, it's a bit faster than a stock -7800X. Stock Threadripper chips land at the bottom of our chart, illustrating the gains available if you're willing to tune AMD's HEDT line-up. The overclocked 1950X also delivers the best 99th percentile results, albeit by a slim margin.

Hitman (2016)

We encountered quite a bit of variability with the Threadripper processors (particularly from the 1950X) during our Hitman test sequence. Even tuning does little to rectify the frame time outliers. Low minimum frame rates, usually misleading when viewed in a vacuum, apply to all the Ryzen models, indicating this title might be particularly sensitive to memory latency.

Middle-earth: Shadow of Mordor

The Core i9-7900X paints a very poor picture of frame time variance as it works through the benchmark, while Threadripper provides a much smoother experience in this older title. AMD's top models also deliver excellent average frame rates.

We observe slight differences between both companies' high-end CPUs as Intel's stock Core i7-7700K dominates.

MORE: Best CPUs

MORE: All CPUs Content

Project CARS & Far Cry Primal

Project CARS

Project CARS is a CPU-intensive title that promotes parallelism by breaking tasks into smaller chunks and spreading them among available cores.

Intel's high IPC throughput is clearly a factor as the Core i7-7700K and i9-7900X cruise to a substantial lead.

Ryzen Threadripper 1920X offers higher clock rates than then 1950X, which helps it secure a narrow victory.

The Core i7-7820X suffers due to its comparatively low core count.

Far Cry Primal

Threadripper 1920X shines when we overclock it. Notably, the chip yields a 14.9ms 99th percentile measurement, while the overclocked -7900X registers a much higher 18.2ms.

The Core i7-7700K suffers several extreme frame time outliers.

MORE: Best CPUs

MORE: All CPUs Content

Rise of the Tomb Raider & The Witcher 3: Wild Hunt

Rise of the Tomb Raider

Rise of the Tomb Raider has long been a thorn in Ryzen's side due to architectural eccentricities. Recent game patches have cleared up most of the inexplicable anomalies, thankfully.

The tuned 1950X fares well, beating our overclocked Threadripper 1920X sample. All of these CPU deliver a smooth experience, though.

We continues to observe that telltale AMD frame rate behavior at the end of the benchmark as it becomes more CPU-intensive, even from the Threadripper models.

The Witcher 3: Wild Hunt

That large frame time spike in the benchmark's early section is a scene transition. All of our contenders suffer from it to some extent, though AMD's Threadripper 1950X stumbles most.

Intel's Core i7-7700K remains the processor to beat.

MORE: Best CPUs

MORE: All CPUs Content

DTP, Office, Multimedia & Compression Performance

Although we usually don’t run our application benchmarks on overclocked processors, we're including the Threadripper models at stock and overclocked frequencies to study how well AMD’s architecture scales with increased frequency. We also added a (reasonably) overclocked Intel Core i9-7900X to our results.

DTP & Presentation

Adobe’s Creative Cloud gives us a look at single- and multi-core performance. As such, it beats synthetic benchmarks as a productivity test.

After Effects CC scales well with core count, granting an easy win to the Threadripper processors. Conversely, InDesign CC finds Intel’s Skylake-X in the back of the pack, while Core i7-7700K leads with its blend of high IPC throughput and frequency. Likewise, AMD’s Ryzen 7 outperforms the Threadripper processors, even after we boost their clock rate. The 1920X dominates in Adobe Illustrator, while the 1950X inexplicably struggles.

Encoding & Multimedia

After tuning, the Threadripper processors trade blows with Intel's overclocked Core i9-7900X. They do lag behind at stock settings; however, the 1920X slips past Core i7-7820X.

Intensifying the workload with higher quality settings really emphasizes threading. This gives AMD's Threadripper processors a sizeable advantage over the -7900X and a resounding win over the rest of the group.

Compression & Decompression

The 1920X's extra cores also come into play during our compression test, giving it an advantage over the stock -7900X.

Intel's overclocked Core i9-7900X barely beats the stock and overclocked 1920X, but its advantage is small enough to be imperceptible.

A lightly-threaded 7-Zip decompression workload prefers high clock rate and IPC throughput. The 1920X's slight frequency advantage over AMD's brawny flagship translates to a quantifiable win. Tuning widens the gap even further.

MORE: Best CPUs

MORE: All CPUs Content

2D & 3D Workstation Performance

2D Workstation Performance

Our GDI/GDI+ tests are used to test two different output methods that can be found in older applications and printing tasks. Today, they, or at least a modified version of them, are commonly used to display the graphical user interface (GUI). They are also great benchmarks for direct device write throughput and memory performance when handling gigantic device-independent bitmap (DIB) files.

Tom’s Hardware Synthetic 2D Benchmarks

We take a look at direct device write throughput first. There hasn’t been true 2D hardware acceleration since the introduction of the unified shader architecture, and Microsoft's Windows driver model complicates 2D hardware acceleration as well.

The graphics test is lightly threaded, so AMD's Threadripper processors struggle while Intel's Core i7-7700K enjoys a massive advantage.

Next, we generate the graphics output in memory, using the only remaining 2D hardware function. The benchmark is the same as before, but we instead plot a bitmap in memory rather than send the information directly to the monitor. The bitmap is only copied once it's complete. This pushes the CPUs, since they’re no longer platform-bound.

Ryzen Threadripper processors dominate the field, with AMD's 1920X landing in first place.

AutoCAD 2016 (2D)

The AutoCAD 2D benchmark doesn't scale well with additional cores. That shifts the focus to IPC throughput, where Intel's processors shine.

Surprisingly, Ryzen 7 1800X beats the tuned Threadripper 1920X. Die-to-die latency may come into play during this test.

3D Workstation Performance

Most professional development applications have been optimized and compiled with Intel CPUs in mind. This is reflected in their performance numbers. Still, we include them in order to motivate developers to focus their efforts on AMD’s Ryzen processors as well. This would give users more than one choice. The same goes for an emphasis on multi-core processors, at least where that’s feasible and makes sense.

AutoCAD 2016 (3D)

AMD’s Ryzen family lands within a narrow range during this frequency-sensitive application. The Core i7-7700K takes an easy lead, and the -7820X's second-place finish confirms that the workload isn't optimized for parallelism. In fact, AutoCAD’s performance resembles older games because it uses DirectX and doesn't leverage multiple cores effectively.

Cinebench R15 OpenGL

Clock rate tends to influence the Cinebench R15 OpenGL benchmark results most. Our numbers indicate that the application could benefit from Ryzen-specific optimizations.

SolidWorks 2015

SolidWorks 2015 tells a similar tale. Even an overclocked Ryzen Threadripper 1920X loses to Ryzen 7 1800X. Switching to NUMA mode could help improve Threadripper's placement, but that'd also have an impact on other applications. You'll have to choose the settings that yield the best experience or face a steady stream of reboots to optimize AMD's platform for whatever you're running at the moment.

Creo 3.0

The 1920X's lead over the 1950X tells us that frequency is more important than core count during this benchmark.

Moreover, Ryzen 7 1800X's performance advantage over the Threadripper processors suggests the unique MCM design can be problematic in some workloads. Optimized BIOS settings could push those processors up in our field, but they'd also negatively impact Creo's CPU composite score.

Blender (Real-time 3D Preview)

Threadripper's Blender benchmark results are acceptable for a high-end processor, but the -7700K is a fly in the ointment for both Intel's and AMD's priciest chips. The overclocked 1920X leads AMD's line-up, illustrating the advantages of extra clock rate in workloads not well-optimized for high core counts. Given the great rendering results you'll see on the next page, Threadripper 1920X provides a potent balance of performance in all types of applications.

Catia V6 R2012

This graphics benchmark is well-optimized (it’s part of the free SPECviewperf 12 suite, after all). We can see that frequency is particularly important in determining performance.

Maya 2013

Maya 2013 also leans heavily on clock rate. But you have to remember that real-time 3D output benchmarks don’t tell the whole story. AMD’s Threadripper processors are much more competitive during final rendering, which we'll cover on the next page.

MORE: Best CPUs

MORE: All CPUs Content

CPU Rendering, Scientific & Engineering Computations, & HPC Performance

CPU Workstation Performance

The 3D graphics performance we just measured isn’t all that matters to professional rendering software. Applications run many other tasks (like simulations, compute jobs, preview rendering) on the CPU simultaneously. The full picture’s only achievable by looking at both of them together.

Many modern suites include modules that are based exclusively on computing and simulations. This means we need to go beyond just 3D workstation performance to form our opinion of these high-end CPUs.

Threadripper offers solid performance in applications that respond well to IPC throughput and frequency, so the Threadripper 1920X leads its counterpart in the SolidWorks and Creo Composite tests. 3ds Max 2015's CPU Computing test benefits from both frequency and parallelism, rewarding the 1950X with a first-place finish when it's overclocked.

CPU Performance: Photorealistic Rendering

Final rendering doesn’t require a CPU that's good at everything. Rather, this task wants efficiency and fast parallel computation.

Nothing beats AMD’s Ryzen Threadripper when it comes to rendering in 3ds Max 2015. Core count is much more important than clock rate, which yields a predictable pecking order.

Most of the Blender tests naturally respond well to Threadripper's array of execution resources.

IPC throughput and clock rate become a factor during the "Blenderalla" workload, so Intel's Skylake-X processors assume the lead.

Overall, the Threadripper products are great for semi-professional use and rendering workloads. Even when they don't win absolutely, they still offer performance that's competitive with Intel's similarly-priced models.

Scientific & Engineering Computations, & HPC Performance

For these tests, we use the SPECwpc benchmark suite for workstations with its wide variety of tasks. It runs a number of different mathematical computations optimized for parallelization. They typically make heavy use of available memory bandwidth and cache, plus expose issues with latency.

The convolution benchmark consists of an operation performed on two functions that results in a third function. Performance scales similarly well with core count and clock rate, allowing the Threadripper architecture to shine.

CalculiX is based on the finite element method for three-dimensional structural computations. In this test, Ryzen Threadripper 1950X bests Intel's -7900X, while the 1920X leads the rest of the field.

SRMP algorithms are used for discrete energy minimization, and none of the AMD entrants handle them well, suggesting the benchmark might be latency-sensitive.

The Kirchhoff Migration test plays well with AMD's Threadripper design, though, so the company's highest-end chips lead the field by an impressive margin.

MORE: Best CPUs

MORE: All CPUs Content

Overclocking, Cooling & Temperature

The Right Cooling Solution

AMD uses solder between Threadripper's dies and its heat spreader. This plays a critical role in aiding our overclocking efforts, helping us achieve the highest clock rate we've seen from a Ryzen CPU. We made significant changes to our cooling setup to avoid a few issues with the standard mounting mechanism.

Overclocking

Our Threadripper 1950X sample overclocked to 3.9 GHz at 1.35V, and the 1920X hit 4.1 GHz with 1.42V. The all-in-one water-cooler in our kit couldn't keep the CPU stable at that level, though. This was due to processor power consumption exceeding 250W during our rendering test.

And so we went back to our higher-end Chiller for a better comparison data. This way, there’s one true constant to our measurements: a water temperature of approximately 20°C that can be held constant, even subjected to more than 300W of waste heat.

For everyday use, a normal water-cooling solution will definitely suffice thanks to AMD's soldered heat spreader, which makes Threadripper less of a challenge than Intel's Core i9-7900X. Our approach simply allows us a bit more thermal headroom.

Using the Chiller, AMD’s Ryzen Threadripper 1920X went even higher, reaching 4.2 GHz at 1.45V and 4.3 GHz at 1.5V. At that point, the Chiller started falling behind Threadripper's thermal output as temperatures crept too high for comfort. We dialed in a more reasonable 4.1 GHz for our benchmarks.

Maximum Temperatures: Stock Clock Rate

For our baseline results, we used the 360mm closed-loop cooler that AMD provided. As a spoiler, the motherboard limits both Threadripper chips to ~180W. You can’t exceed this upper boundary for even short periods using normal settings.

The 1920X’s curves look almost identical to those of AMD's Ryzen Threadripper 1950X, except for a few insignificant Tctl and Tdie jumps. This is hardly surprising, seeing that the load is similar as well.

The CPU temperature values reported by HWiNFO64 through Asus' separate sensor loop are between 6°C and 12°C lower than the Tctl values, and they rise more slowly. Even without additional cooling, the voltage converter temperatures are almost perfect at just under 60°C.

Maximum Temperatures: Overclocked

Increasing voltages to guarantee stable operation pushes the processors past their sweet spot. Consequently, power consumption goes through the roof. Operating well beyond 300W poses a challenge for any cooling solution. That's why we're using the Chiller. We did try a normal water-cooling loop though, resulting in the Tctl and Tdie values going up by ~10°C to 15°C. This is well within an acceptable range.

The overclocked 1950X peaks at 320 to 325W. Using the Chiller, this level of power consumption is accompanied by Tctl values of 87°C. That’s actually not as severe as it seems once the offset and Tdie values are taken into account. A real temperature of approximately 60°C serves as a great demonstration of why solder is superior to thermal paste.

MORE: Best CPUs

MORE: All CPUs Content

Power Consumption

We establish the package’s power consumption results by using a special sensor loop. This way, our values represent the exact amount of power that goes into the CPU and then reemerges in the form of waste heat dissipated by the cooling subsystem. We check our sensor readings using shunts and by measuring overall power consumption directly at the EPS connector (with a current probe and direct voltage measurement).

AMD’s Threadripper CPUs use different partial voltages for the SoC and SMU rails at different clock rates. These partial voltages, which, again, vary based on frequency, do influence the package’s power consumption. AMD recommended that we use the profile included with its DDR4-3200 kit. But if we instead use the standard SPD values for DDR4-2133, our power measurement is 15W lower!Both of AMD’s CPUs are designed for a maximum power ceiling of 180W at their default settings. If the memory gets overclocked, the CPU has 15 fewer watts to work with. This could affect performance in workloads that utilize all cores and, consequently, get too close to the limit.

Idle Power Consumption

Threadripper’s idle power consumption is roughly twice that of the Ryzen 7 models. However, Threadripper also hosts two dies instead of one, and it also hits higher clock rates under sporadic loads. The overclocked version requires higher voltages as well, and memory also plays a role in power consumption. For instance, dropping to DDR4-2133 pulls the 1920X's idle power use down to 32W.

CAD Workload Power Consumption

AutoCAD 2016 rarely uses more than two or three cores. In fact, most of the time it's limited to a single core. Thus, it's not surprising that the CAD power consumption only adds a maximum of 15W to the idle power numbers. The two overclocked configurations add another 14W, which makes for an almost 30W difference compared to our idle power consumption results.

Gaming Power Consumption

When it comes to gaming, Threadripper’s MCM design causes its many cores to get in each others' way. Thus, the frame rates we report end up lower than competing processors. But power consumption ends up similar to Intel's Core i9-7900X, even though Skylake-X offers much more performance.

Stress Test & Maximum Power Consumption

Power consumption goes through the roof during our stress test, especially for the overclocked configurations.

The motherboard is partially to blame for the stock Intel Core i9-7900X's excessively high numbers. It doesn’t obey the standard Turbo Boost frequency thresholds, instead boosting aggressively and staying in those boost states longer than required. For more details, see our article about the power and thermal issues we encountered during our extended testing.

Threadripper doesn’t have those kinds of issues. Asus X399 ROG Zenith Extreme limits power consumption to exactly 180W at stock settings, just as it should.

At a respectable 1.425V, the Ryzen Threadripper 1920X reaches 4.1 GHz. The higher-end 1950X needs 1.35V to achieve 3.9 GHz. Once overclocked, AMD’s new processors join Intel's Core i9-7900X overclocked to 4.5 GHz in the stratosphere beyond 300W.

In the end, Threadripper's two dies sometimes consume more power than other processors’ single dies, depending on the task. We succeeded in breaking the 4 GHz barrier by overclocking the 1920X to 4.1 GHz. At that speed, all 24 threads were fully functional and at our disposal. The high power consumption is acceptable if it's accompanied by comparably elevated application performance. For Threadripper, that requires highly parallelized workloads (and perhaps optimized software).

Unfortunately, Threadripper's efficiency during gaming turns out to be significantly worse than Intel’s. Threadripper draws an additional ~15W at idle due to the memory. Subtracting that 15W from AMD's gaming power consumption changes the picture, bringing power consumption in line with the lower gaming performance.

MORE: Best CPUs

MORE: All CPUs Content

Final Analysis

Ryzen Threadripper 1920X bears most of the same features touted on Threadripper 1950X. It clearly offers strong performance in threaded applications, but it also comes with higher base clock rates and more overclocking headroom than any Ryzen model we've tested. Compared to the 1950X, you save $200 in exchange for four cores and eight threads. However, you also gain higher performance in many lightly threaded productivity applications.

AMD positions Threadripper as a solution for content creators, heavy multi-taskers, and gamers who stream to services like Twitch. It also says the processors are ideal for gaming at high resolutions (a most logical pairing, given the likely specs of a desktop with an $800 CPU). The 1920X isn't intended for low-resolution gaming, particularly with lightly threaded titles. Still, we test at lower resolutions to unearth the differences between competing architectures, rather than be bound by graphics performance.

The following gaming price efficiency charts use a geometric mean of the 99^th percentile frame times (a good indicator of smoothness), which we convert into an FPS measurement and plot against price. Our suite includes six games released in 2016 and five older titles that launched in 2014/2015. Threadripper’s extra cores could enable more performance in the future as software evolves to utilize them better, so we also include a chart with newer games that exploit host processing resources more thoroughly.

Ryzen Threadripper 1920X drops into the gap between Intel's $600 Core i7-7820X and $1000 i9-7900X. It offers less performance than the Intel processors in both new and older games, even after a substantial overclock. Those deltas will shrink at higher resolutions, though. The 1920X's performance is fairly comparable to the higher-end 1950X, although AMD's flagship also exhibits a relatively small lead over the 1920X in stock and overclocked configurations.

Threadripper's true value registers in more intense workloads, such as heavy multitasking while gaming. Moreover, its hefty allotment of 60 available PCIe lanes allows for plenty of expansion. Even though the X399 motherboards are quite stable, more performance-enhancing firmware is trickling out from several vendors. We've already seen much higher gaming performance from the 1950X in Game mode, which is promising. Ryzen-specific optimizations for current titles continue surfacing as well, and we expect most new games to include similar optimizations. Gaming on Ryzen should only improve with time.

Of course, we still recommend sticking with mainstream processors like Ryzen 7/5 or Core i7/i5 for the best gaming value. That recommendation applies to both Intel and AMD high-end CPUs.

Focusing more on Threadripper's core competency, the 1920X offers great performance in a few of our less demanding productivity tasks, such as the Adobe suite. Notably, the 1920X's extremely high score in Adobe Illustrator feels like an outlier, so we provide charts both with and without that test. In either case, the 1920X's frequency advantage provides more performance than Core i9-7900X in this and some other lightly threaded tasks, like decompression.

The 1920X excels in encoding and compression workloads, often matching or outstripping Intel's Core i9-7900X. The 1920X isn't as dominant in the Blender and LuxRender tests, but it delivers incredibly competitive performance, especially in light of its lower price point. It also fares well in many of our HPC and scientific workloads, highlighting its diverse capabilities.

The Threadripper processors are a solid choice for highly parallelized or simultaneous workloads. Intel still enjoys an advantage in most lightly threaded tasks. But overall, the 1920X is more competitive in these applications than the lower-frequency 1950X. Of course, switching into Game mode might enable higher performance in some situations, but we don't think professional users will tolerate constant reboots to toggle back and forth.

Intel's X299 and AMD's X399 platform costs are similar, at least by early indications. Several TR4-specific coolers have already come to market, and we expect more in the future. Surprisingly, the bundled Asetek bracket, which provides poor IHS coverage, is sufficient to attain substantial overclocks (at least by Ryzen standards). We used the bracket and a standard Thermaltake 360mm radiator to achieve a rock-solid 4.1 GHz, so cooling isn't as much of a worry here as it was with Skylake-X. Take note Intel; solder pays off.

Intel's Skylake-X models are still trickling out, so the company will have faster options soon. But they'll launch at hideous price points. Meanwhile, the 1920X slots into the $400 chasm between Core i9-7900X and i7-7820X, and it doesn't appear that Intel will have a Skylake-X processor to compete any time soon. This is a tremendous opportunity for AMD, and it's great news for anyone seeking no-compromise connectivity, competitive responsiveness in everyday apps, and superior performance per dollar in threaded software.

MORE: Best CPUs

MORE: All CPUs Content

AMD Ryzen Threadripper 1950X Review

palcorn@outlook.com (Paul Alcorn) — Thu, 10 Aug 2017 13:11:00 +0000

Threadripper Makes An Entrance

AMD's 16C/32T Ryzen Threadripper is likely the biggest processor release of the year, quite literally. And that's saying a lot, given the company's rapid-fire releases addressing almost every facet of the desktop market. Now AMD is attacking Intel's high-end desktop stronghold with up to 60% more cores, 36% more PCIe lanes, and 68% more cache than the comparably-priced Skylake-X models. AMD also doesn't hack and slash at its product stack by culling I/O, so the company offers the same connectivity, even on its least-expensive models.

Impressively, all of the new Ryzen 7, 5, and 3 processors utilize the same underlying modular architecture. Threadripper is no exception, though it borrows some of AMD's EPYC data center DNA to take the design to the next level.

Four years ago, Dr. Lisa Su, now AMD's CEO, tasked her team with building a server processor that could compete with Intel's finest. That was an ambitious goal considering how far the company lagged behind its main competitor. AMD's Zen core, which is a clean-sheet design that provides a 52% IPC boost over the company's previous processors, was already underway and destined to power the new chips.

But processor design is a series of compromises. The company quickly learned that it couldn't construct a single monolithic die able to deliver on its performance, memory, and I/O goals. Instead, AMD's engineers put together modular four-core building blocks (CCXes) married in an eight-core die. And so, Zeppelin was born.

Now AMD uses the same Zeppelin building block in all of its processors. This approach is the epitome of maximizing limited resources; AMD can simply add more dies per package to create massively parallel chips, such as Threadripper and EPYC. Of course, there are trade-offs...

Though AMD faced significant headwinds at launch due to its unique Ryzen design, the complementary platform is much more mature, and game developers are doing better about leveraging the architecture's potential. AMD's efforts are paying off, too: Ryzen CPUs continue taking over our list of Best CPUs. Threadripper brings an entirely new set of challenges, but AMD has already designed in a few new features to help on the performance front.

Threadripper 1950X, 1920X & 1900X

It's important to understand Threadripper's intended audience. AMD geared the design for software developers, video/audio engineers, content creators, and heavy multi-taskers. Though Threadripper isn't aimed directly at gamers, particularly those who play at low resolutions with lightly-threaded titles, the company makes provisions for entertainment. The hefty core count should also boost gaming performance during intense streaming workloads, which we are currently testing for an upcoming feature.

The $1000 Ryzen Threadripper 1950X is the leader of AMD's high-end desktop portfolio with 16C/32T, while the smaller 12C/24T 1920X and 8C/16T 1900X round out the family. Much like the rest of its Ryzen line-up, AMD's Threadripper processors offer more cores than Intel at every price point. This time, however, we also get 64 PCIe 3.0 lanes (four of which are dedicated to the chipset) that outweigh Intel's 44-lane flagship Core i9-7980XE. And that chip isn't even available yet.

AMD includes all 64 lanes with each of its three Threadripper models, whereas Intel reduces connectivity for Skylake-X processors below $1000. Dual graphics cards are becoming more of a rarity in high-end gaming machines, but there are plenty of workloads still constrained by I/O. For instance, many taxing content creation workloads require hefty storage accommodations, and streamers often employ dedicated capture cards.

	Threadripper 1950X	Core i9-7900X	Threadripper 1920X	Core i7-7820X	Threadripper 1900X
Price	$999	$999	$799	$599	$549
Interface/Chipset	TR4 / X399	LGA2066 / X299	TR4 / X399	LGA2066 / X299	TR4 / X399
Cores/Threads	16/32	10/20	12/24	8/16	8/16
TDP	180W	140W	180W	140W	180W
Base Frequency (GHz)	3.4	3.3	3.5	3.6	3.8
Boost Frequency (GHz)	4.0 (4.2 XFR)	4.3 / 4.5 (TB 3.0)	4.0 (4.2 XFR)	4.3 / 4.5 (TB 3.0)	4.0 (4.2 XFR)
L3 Cache (L2+L3)	40 MB	23.75 MB	38 MB	19MB	20 MB
Memory Support	DDR4-2667	DDR4-2666	DDR4-2667	DDR4-2666	DDR4-2667
Memory Controller	Quad-Channel	Quad-Channel	Quad-Channel	Quad-Channel	Quad-Channel
Unlocked Multiplier	Yes	Yes	Yes	Yes	Yes
PCIe Lanes	64	44	64	28	64

The Threadripper 1950X offers a 3.4 GHz base frequency that boosts to 3.6 GHz during heavily-threaded workloads. Mainstream Ryzen models feature a dual-core Precision Boost during lightly-threaded tasks, but due to Threadripper's dual-die design (which we'll cover on the following page), that expands to a quad-core 4.2 GHz boost. Much like Intel's Skylake-X series, AMD offers unlocked multipliers for all Threadripper models, but improves overclocking frequency and voltage scaling by selecting the top 5% of Zeppelin dies. That should equate to lower voltage requirements compared to Ryzen 7 models.

Each Zeppelin die also sports 16MB of L3 cache, so the Threadripper 1950X wields a total of 32MB. That's a lot more than the 24.75MB Intel enables on its upcoming 18C/36T Core i9-7980XE. Focusing in closer on the $1000 price range, the 1950X provides 18.25MB more L3 cache than Core i9-7900X. Of course, cache latency and bandwidth can be much more important than capacity, so we'll have to measure each company's balance between size and speed.

AMD's dual-die architecture also comes with a 180W TDP rating, which is comparatively higher than Skylake-X's 165W ceiling. Of course, TDP doesn't map directly to power consumption during all workloads. AMD has an extensive set of power reduction features in its SenseMI suite that should help, and we'll present plenty our findings shortly.

Of course, a large 180W TDP processor is going to require a beefy thermal solution, but there aren't any water-coolers designed specifically for Threadripper's massive heat spreader yet. To that end, AMD includes an Asetek bracket to provide compatibility with a wide range of existing closed-loop coolers. The current solution is an adequate stopgap until custom TR4 solutions arrive, but coverage isn't optimal. There are also several air-cooling options available, including special Noctua TR4 designs with larger bases that cover the entire heat spreader surface.

Threadripper features independent dual-channel memory controllers, one per die, that combine to provide quad-channel support with varying data transfer rates (outlined below) based upon memory configurations. The platform supports ECC memory and a functional limit of 256GB of RAM, though it can support up to 2TB of capacity as memory density increases.

Ryzen Threadripper Memory Support	MT/s
Quad-Channel/Dual-Rank/Two DIMMS per Channel (8)	1866
Quad-Channel/Single-Rank/Two DIMMs Per Channel (8)	2133
Quad-Channel/Dual-Rank/One DIMM Per Channel (4)	2400
Quad-Channel/Single-Rank/One DIMM Per Channel (4)	2677

AMD connects the two Zeppelin die via its Infinity Fabric, so data resident in "remote" memory banks suffers higher latency. The company includes two new settings that help offset disadvantages inherent to the distributed memory architecture, which we'll dive into on the next page. A Creator Mode offers the full heft of compute and memory resources for productivity workloads, while Game Mode tries to keep data resident in "local" memory and restricts processing to a single die. The idea of dedicated modes is certainly new to the desktop, and it requires a reboot after each change, but we found that it does offer tangible performance benefits in gaming workloads.

The X399 chipset also debuts with a new TR4 socket in tow, but it is an enthusiast-class platform that has roots in the data center. That means corresponding motherboards will carry a premium over AMD's AM4 line-up, and may even rival the prices of Intel's X299-equipped boards. Naturally, AMD has a solid roster of motherboard partners ready for its Threadripper launch, thus avoiding the maddening delay we experienced with the first Ryzen processors.

In case you had any doubt, this is a premium product. AMD outdoes Intel's packaging with a custom closed-cell foam design that includes a torque wrench. Head over to our unboxing coverage to see the Threadripper packaging in all its glory.

But we're here for performance testing. Let's see if Threadripper lives up to the hype.

MORE: Best CPUs

MORE: Intel & AMD Processor CPU Benchmarks Hierarchy

MORE: All CPUs Content

Game Modes & Architecture, Infinity Fabric Latency Testing

We've covered AMD's Zen architecture in depth, and also covered the Infinity Fabric at length. Head over to those articles for more coverage.

The Zeppelin Die Primer

Threadripper's massive package hides much complexity underneath, but we'll do our best to simplify and outline how it relates to AMD's innovative Creator and Game Mode features.

The Zen architecture employs a four-core CCX (CPU Complex) building block. AMD adorns each CCX with 8MB of L3 cache split into four slices; each core in the CCX accesses all L3 slices with the same average latency. Two CCXes come together to create an eight-core Ryzen 7 die (the large orange blocks in the second image below), and they communicate via AMD’s Infinity Fabric interconnect. The CCXes share the same dual-channel memory controller. This is basically two quad-core CPUs talking to each other over the Infinity Fabric pathway that also handles northbridge and PCIe traffic.

All Ryzen 7, 5, and 3 models feature the same single Zeppelin die. Although each core in a four-core CCX can access the local cache with the same average latency, trips to fetch data in adjacent CCXes incurs a latency penalty. Communication between threads on cores located in disparate CCXes also suffers, which is of particular importance for gaming. Many game engines split out various tasks to different threads, but they are reliant upon constant synchronization between them. Developers can defray some of the communication latency by tuning for the Ryzen architecture.

Building The Threadripper

The graphic below represents AMD's EPYC data center processor die, which shares Threadripper's basic design. We can see four separate Zeppelin dies connected via the Infinity fabric, and the two CCXes inside each die. This creates a 32-core Multi-Chip Module (MCM). Of course, Threadripper is "only" a 16-core processor. To create this configuration, AMD substitutes in two 'dummy dies,' which are non-functional fillers that ensure the heat spreader's structural integrity and consistent mating with the socket's pins. Without these dark dies, the IHS would either cave in when you tighten your cooling solution, or the chip would warp and not make full contact with the pins. AMD notes that Threadripper's functional dies are always placed diagonally from each other, which makes sense considering the fabric's design.

Remember, each Zeppelin die has its own memory and PCIe controllers. That means that if a workload executing on a die needs to access data resident in the memory of the other die (remote memory), it has to traverse a much larger gap. This introduces a level of latency we haven't seen from previous Ryzen models, and its effect on gaming performance is profound. The impact isn't as severe with most professional workloads, but some do suffer.

The New Toggles

To defray the impact of remote memory access, AMD introduces a new memory access mode that you can toggle either in the BIOS or with the Ryzen Master software. The Local and Distributed settings flip between either NUMA (Non-Uniform Memory Access) or UMA (Universal Memory Access).

UMA (distributed) is pretty simple; it allows the dies to access all of the attached memory. NUMA mode (local) attempts to keep all data for the process executing on the die confined to its directly attached memory controller. It establishes one NUMA node per die (visible in the task manager). This reduces, and even possibly eliminates, data fetches from the remote memory connected to another die, though the die can still access it if needed. NUMA has deep roots in the enterprise, but the technique works best if programs are designed specifically to utilize it. It's a rarity on the desktop, but even though almost no desktop applications are designed to support it entirely, there can be performance advantages for non-NUMA applications.

AMD's Threadripper introduces more cores to the desktop than we've ever seen; some programs are caught ill-prepared. In fact, a few games like Far Cry Primal and the DiRT series won't even run when the full complement of Threadripper's threads are brought to bear. That's obviously a problem, so AMD created a Legacy Compatibility mode that disables half of the processor's cores by executing a "bcdedit /set numproc XX" command in Windows that effectively disables half of the processor. Luckily, due to the operating system's core assignments, the command disables all of the cores/threads on the second die. That has a side benefit of eliminating thread-to-thread communication between disparate die, serving as a great solution to the constant synchronization between threads during most gaming workloads.

Because the change is made in software, the "disabled" die still has power fed to it, so the system can still access the memory and PCIe controllers connected to the inactive die.

Game Mode And Creator Mode

So what do you do with all these knobs? There are four separate combinations that will impact each application or game differently, so you have to cycle through them to find the best possible combination for your workload. That's a godsend to tuners looking to squeeze out every last drop of performance, but an absolute nightmare for the other 99%.

AMD decided to simplify the process by specifying two combinations that will either work best for games or standard applications. Creator mode, which is the stock configuration, exposes the full might of 32 threads. It should naturally provide excellent performance for most productivity applications.

Game mode cuts half the threads via compatibility mode and reduces memory and die-to-die latency with the Local memory mode. We're going to test both configurations with our gaming suite, and try another configuration that also offers the full complement of threads.

Infinity Fabric Latency Testing

Die-to-die communication adds another layer of latency to Ryzen’s complicated architecture. As you can see, those same latency metrics don’t apply to the earlier Ryzen models. They also present challenges to some applications, such as those with synchronized threads or frequent fetches from remote memory, but have less impact on others.

Processor	Intra-Core Latency	Intra-CCX Core-to-Core Latency	Cross-CCX Core-to-Core Latency	Cross-CCX Average Latency	Die-to-Die Latency	Die-To-Die Average Latency	Average Transfer Bandwidth
TR 1950X Creator Mode DDR-2666	13.7 - 14.1	39.4 - 43.2ns	157.6 - 171.3	168ns	180.6 - 256.7ns	238.47ns	90.26 GB/s
TR 1950X Creator Mode DDR4-3200	13.8 - 14.9	39.2 - 45.4ns	144.9 - 167.2ns	160.1ns	213.1 - 227.8ns	216.9ns	91.67 GB/s
TR 1950X Game Mode DDR4-2666	13.9 - 14.2ns	39.5 - 42.3ns	149.2 - 164.1ns	159.66ns	X	X	46.58 GB/s
TR 1950X Game Mode DDR4-3200	14.3 - 14.9ns	41.2 - 46.2ns	123 - 150.6ns	145.44ns	X	X	45.52 GB/s
TR 1950X Local/SMT DDR4-2666	13.9 - 14.4ns	39.6 - 43.1ns	168.7 - 175.4ns	171.48ns	232.4 - 240.8	235.38ns	92.7 GB/s
TR 1950X Local/SMT DDR4-3200	13.9 - 14.4ns	39.9 - 44.5ns	146.7 - 159.4ns	153.89ns	209.3 - 220.9ns	212.53ns	91 GB/s
Ryzen 7 1800X	14.8ns	40.5 - 82.8ns	120.9 - 126.2ns	122.96ns	X	X	48.1 GB/s
Ryzen 5 1600X	14.7 - 14.8ns	40.6 - 82.8ns	121.5 - 128.2ns	123.48ns	X	X	43.88 GB/s

The intra-core latency measurements represent communication between two logical threads resident on the same physical core, and they're unaffected by memory speed. Intra-CCX measurements quantify latency between threads that are on the same CCX but not resident on the same core. In the past, we observed slight performance variances, but intra-CCX latency is also largely unaffected by memory speed. However, we've seen a large decrease in cross-CCX latency, which denotes latency between threads located on two separate CCXes, by increasing the memory data transfer rate from DDR4-1333 to DDR4-3200 on Ryzen 5 and 7 models.

The same general trend continues with Threadripper. As we can see, toggling game mode removes the die-to-die latency for threads by effectively disabling one die, but it also reduces host processing resources. It’s an interesting feature that will benefit some workloads, but hamstring others.

We also notice that the Local/SMT combination, which consists of the local setting and leaves all cores active (legacy off), offers the best overall latency improvement via memory overclocking. We also recorded higher Cross-CCX latency with the Threadripper processors.

Processor	Intra-Core Latency	Core-To-Core Latency	Core-To-Core Average Latency	Average Transfer Bandwidth
Core i9-7900X	14.5 - 16ns	69.3 - 82.3ns	75.56ns	83.21 GB/s
Core i9-7900X @ 3200 MT/s	16 - 16.1ns	76.8 - 91.3ns	83.93ns	87.31 GB/s
Core i7-6950X	13.5 - 15.4ns	54.5 - 70.3ns	64.64ns	65.67 GB/s
Core i7-7700K	14.7 - 14.9ns	36.8 - 45.1ns	42.63ns	35.84 GB/s

We are in the midst of a broader set of tests to quantify how these modes impact memory latency and bandwidth, among other factors. Stay tuned.

MORE: Best CPUs

MORE: All CPUs Content

TR4 Socket, X399 Chipset & Test Setup

The Threadripper TR4 Socket

Accommodating AMD's massive Threadripper processors required a new interface. Socket TR4 (SP3r2) for X399 motherboards boasts 4094 pins, which is much more complex than Intel's LGA 2066 interface found on X299-equipped motherboards. Socket TR4 socket marks AMD's transition from a PGA (Pin Grid Array) with the first Ryzen models to an LGA (Land Grid Array) implementation.

The processor employs a unique mounting process that should be safer than the normal latching mechanism, all the while remaining simple. Just remove the three Torx screws in a pre-defined order and the retention mechanism swings open. There is another assembly underneath that you also swing open. Pull out an internal plastic cover, and then slide the processor—contained in an orange plastic shroud—into the carriage. It's important to leave the black cover over the socket until you install the processor in the flip-up housing. Those pins are fragile.

After sliding the processor into its carriage, you remove the socket cover and swing the processor down into the socket. Then you swing down the mounting mechanism and tighten the fasteners (in the listed order) with the bundled Torx T20 torque wrench. We've included a picture of AMD's recommended TIM application technique. Heat sink installation requires four additional fasteners.

Socket AM4, like Threadripper, has its roots in the EPYC data center design. But the processors aren't interchangeable.

The X399 Chipset

The X399 chipset supports two USB 3.1 Gen2 and six USB 3.1 Gen1 ports, along with six USB 2.0 connections. Two PCIe 3.0 lanes allow motherboard vendors to add more storage connectivity (four SATA or two SATA Express), and the eight general-purpose PCIe 2.0 lanes accommodate other controllers, such as Ethernet or WLAN/Bluetooth. Eight SATA ports round out the chipset's connectivity options, and you can leverage several RAID configurations with the attached SATA devices. Unfortunately, NVMe RAID isn't supported in hardware yet, though software RAID is still an option.

Comparison Products

Ryzen 7 1800X

Intel Core i9-7900X

Intel Core i7-7820X Skylake-X

Test Systems

We introduced our new test system and methodology in How We Test Graphics Cards. If you'd like more detail about our general approach, check that piece out.

In this case, only the hardware configuration with CPU, RAM, mainboard, as well as the new cooling system are different, so the summary in table form gives a quick overview of the systems used:

Test System and Configuration
Hardware	Germany AMD Socket SP3 (TR4)AMD Ryzen Threadripper 1950XAsus X399 ROG Zenith Extreme4x 8GB G.Skill RipJaws V DDR4-3200Intel LGA 2066Intel Core i9-7900XMSI X299 Gaming Pro Carbon AC4x 4GB G.Skill RipJaws IV DDR4-2600AMD Socket AM4 WorkstationAMD Ryzen 7 1800X, 1700X, 1600X, 1500XMSI X370 Tomahawk4x 8GB G.Skill TridentZ DDR4-3200Intel LGA 2011v3Intel Core i7-6900KMSI X99S XPower Gaming Titanium4x 4GB Crucial Ballistix DDR4-2400Intel LGA 1151Intel Core i7-7700KMSI Z270 Gaming 72x 8GB Corsair Vengeance DDR4-3200 @ 2400 MT/sAll SystemsGeForce GTX 1080 Founders EditionNvidia Quadro P6000 (Workstation)1x 1TB Toshiba OCZ RD400 (M.2, System)2x 960GB Toshiba OCZ TR150 (Storage, Images)be quiet! Dark Power Pro 11, 850W Power SupplyWindows 10 Pro (Creators Update)USAMD Socket SP3 (TR4)AMD Ryzen Threadripper 1950XAsus X399 ROG Zenith Extreme4x 8GB G.Skill RipJaws V DDR4-3200 @ 2666 and 3200 MT/sIntel LGA 2066Intel Core i9-7900X, i7-7820XMSI X299 Gaming Pro Carbon AC4x 8GB G.Skill RipJaws V DDR4-3200 @ 2666 and 3200 MT/sAMD Socket AM4 AMD Ryzen 7 1800XMSI X370 Xpower Gaming Titanium2x 8GB G.Skill RipJaws V DDR4-3200 @ 3200 MT/sIntel LGA 1151 Intel Core i5-7700K MSI Z270 Gaming M72x 8GB G.Skill RipJaws V DDR4-3200 @ 2666 and 3200 MT/sAll EVGA GeForce GTX 1080 FE 1TB Samsung PM863 SilverStone ST1500, 1500W Windows 10 Creators Update Version 1703
Cooling	GermanyAlphacool Eiszeit 2000 ChillerAlphacool Eisblock XPXThermal Grizzly Kryonaut (For Cooler Switch)
Monitor	Eizo EV3237-BK
PC Case	Lian Li PC-T70 with Extension Kit and Mods Configurations: Open Benchtable, Closed Case
Power Consumption Measurement	Contact-free DC Measurement at PCIe Slot (Using a Riser Card) Contact-free DC Measurement at External Auxiliary Power Supply Cable Direct Voltage Measurement at Power Supply 2x Rohde & Schwarz HMO 3054, 500MHz Digital Multi-Channel Oscilloscope with Storage Function4x Rohde & Schwarz HZO50 Current Probe (1mA - 30A, 100kHz, DC) 4x Rohde & Schwarz HZ355 (10:1 Probes, 500MHz) 1x Rohde & Schwarz HMC 8012 Digital Multimeter with Storage Function
Thermal Measurement	1x Optris PI640 80Hz Infrared Camera + PI Connect Real-Time Infrared Monitoring and Recording
Acoustic Measurement	NTI Audio M2211 (with Calibration File, Low Cut at 50Hz) Steinberg UR12 (with Phantom Power for Microphones)Creative X7, Smaart v.7 Custom-Made Proprietary Measurement Chamber, 3.5 x 1.8 x 2.2m (L x D x H) Perpendicular to Center of Noise Source(s), Measurement Distance of 50cm Noise Level in dB(A) (Slow), Real-time Frequency Analyzer (RTA) Graphical Frequency Spectrum of Noise

MORE: Best CPUs

MORE: All CPUs Content

VRMark, 3DMark & AotS: Escalation

Testing Notes

As discussed on page two, Threadripper offers both Creator and Game modes to tailor the processor for different applications. We tested with both to compare the gaming-optimized and out-of-box configurations. Threadripper processors also feature two switches that allow for even more unique configurations, so we also tested with the full complement of cores (legacy off) and memory access set to "local." These “Local/SMT” results provide a nice boost for heavily-threaded games, particularly in overclocked configurations.

Most charts are presented in their entirety. But due to space constraints with our frame time variance graphs, we often include secondary charts with additional Threadripper performance data.

AMD notes that some games prefer different configurations, but we don’t have a detailed list of preferred settings for individual titles. Hopefully our results serve as a guide to gamers looking for the best settings.

Intel also released new microcode for its Skylake-X processors recently, which reduces performance in some titles and lowers the AVX offset by two bins. We also noticed far lower Turbo Boost activation thresholds, though that could separately be the result of MSI's newest BIOS. The changes likely come in response to some of the power and thermal issues we encountered during our extended testing. We consequently retested both Skylake-X processors with the newest microcode.

AMD's Ryzen gaming performance is also a moving target, though it continues to improve over time. Today's story reflects all processors re-tested with the latest chipset, BIOS, GPU drivers, and game patches. We're also deploying a new test image and game suite, so the results contained herein are only directly comparable to our Ryzen 3 1300X review.

VRMark & 3DMark

We aren't big fans of using synthetic benchmarks to measure game performance, but 3DMark's DX11 and DX12 CPU tests provide useful insight into the amount of horsepower available to game engines.

VRMark prizes IPC throughput and frequency. AMD's Ryzen Threadripper processors fall to the bottom of the chart, while the Intel models lead. We suspect that VRMark lacks some of the optimizations incorporated into other games.

The Threadripper 1950X fares much better during the heavily-threaded DX11 physics test, where it leads by a large margin. And it squeezes by the overclocked Core i9-7900X during the DX12 CPU test due to its copious core count. The Core i7-7700K suffers through these two tests due to its quad-core design.

Gaming performance, from a host processing perspective, relies on generating the maximum amount of draw calls (sometimes tens of millions per second) using the relevant API. AMD's Threadripper 1950X shows well in our heavily-threaded DX12 tests. But Vulkan performance suffers with all 32 threads active.

Ashes of the Singularity: Escalation

Ashes of the Singularity: Escalation is a computationally intense title that scales well with thread count. That’s music to Threadripper’s ears as the overclocked configuration leads. Intel's 10-core -7900X puts up a stout defense with 60% fewer cores, though.

The 1950X doesn’t fare as well out of the box; an eight-core i7-7820X manages to beat the 16-core contender.

MORE: Best CPUs

MORE: All CPUs Content

Civilization VI, Battlefield 1 & Dawn of War III

Civilization VI AI Test

Quad-core processors tend to provide the best results in Civilization VI's AI test, which measures available computational horsepower during a turn-based strategy gaming session.

The Core i9-7900X puts on an impressive show, though, as it vies for supremacy with Intel's -7700K. The overclocked 1950X in Game Mode provides solid performance, though it drops to the bottom of our chart in its stock configuration. Interestingly, the 1950X using Local/SMT settings yields the best results in an out-of-the-box arrangement, though it drops down the chart after overclocking. This is the only game in which we see this repeatable trend.

Civilization VI Graphics Test

Skylake-X has a tendency to perform poorly in this test, which is likely due to some of the regressions we encountered with Intel’s new mesh architecture. The stock Threadripper processors outperform Intel’s finest. Notably, the overclocked 1950X with Local/SMT settings also beats the Core i9. You can clearly see a few of the -7900X’s hiccups in our frame time chart. All the while, a tuned Core i7-7700K dominates.

Battlefield 1 (DX11)

The stock Threadripper processors stumble during the first seconds of our test sequence, resulting in much of the frame time variance you see in our chart. The tuned 1950X in Game Mode avoids this early choppiness and provides the best overall performance from AMD. It’s notable that the delta between most of the processors is smaller than we see in other titles, though the overclocked Core i9-7900X provides a bit of extra horsepower to distance itself slightly.

Warhammer 40,000: Dawn of War III

The eight-core i7-7820X struggles during our benchmark run, though it does beat AMD's stock 1950X in Local/SMT mode. We realize a large improvement from Threadripper simply by toggling into Game Mode. Meanwhile, a decent 3.9 GHz overclock doesn’t provide much extra performance. That stands in contrast to the large overclocking gains extracted from Core i7-7700K and i9-7900X, implying this game might benefit from optimization.

MORE: Best CPUs

MORE: All CPUs Content

Grand Theft Auto V, Hitman & Shadow of Mordor

Grand Theft Auto V

We measure performance during Grand Theft Auto V's F-16 flight sequence with the built-in benchmark.

Grand Theft Auto V has always played well with the Intel architecture's high frequencies and IPC throughput, and that trend continues during today's round of testing. Both Core processors respond well to overclocking, while the Threadripper 1950X gains minimal uplift. The overclocked Local/SMT setting enables optimal performance, but it doesn’t significantly affect the finishing order.

The Core i9-7900X provides the best 99^th percentile measurements by a large margin.

Hitman (2016)

We encountered quite a bit of variability with the Ryzen Threadripper processors (particularly from the 1950X) during our Hitman test sequence. Even tuning does little to rectify the frame time outliers. Interestingly, AMD's Game Mode under-performs Creator's Mode by a sizeable margin.

Middle-earth: Shadow of Mordor

The overclocked Core i7-7700K carves out a lead in this lightly-threaded game. The more expensive processors all feature similar average frames rates, while we do encounter several frame time outliers with a tuned -7900X. The Ryzen 7 1800X and Threadripper processors provide the most consistent frame time profiles.

MORE: Best CPUs

MORE: All CPUs Content

Project CARS & Far Cry Primal

Project CARS

Project CARS is a CPU-intensive title that promotes parallelism by breaking tasks into smaller chunks and spreading them among available cores.

This game is dominated by the overclocked Core i7-7700K. Considering its relatively low price point, the Kaby Lake-based chip continued delivering excellent gaming value.

Much like the -7700K outperforms its high-end desktop counterparts, Ryzen 7 1800X also outpaces AMD's pricier CPUs. The Threadripper 1950X beats Intel's -7820X, and tuning widens the gap.

Far Cry Primal

Far Cry Primal does not initialize when it detects Threadripper's 32 threads, so we tested it with Legacy mode selected.

Intel's stock -7700K suffers several disturbing frame time outliers that manifest as visible stuttering during the benchmark.

Ryzen 7 1800X offers almost the same performance as the tuned Threadripper 1950X, suggesting that it might be a better value if you are strictly interested in gaming. The less expensive Ryzen 7 1700 facilitates similar overclocking headroom, and thus performance, in a majority of our game tests. That translates to additional value for enthusiasts willing to overclock.

An overclocked -7700K delivers the best performance by a significant margin during the benchmark's opening seconds. This translates to a higher average frame rate, but a 99th percentile measurement of 15.6ms. The Ryzen 7 1800X and Threadripper 1950X offer slightly better 99th percentile measurements of 15.3ms and 15.1ms, respectively.

MORE: Best CPUs

MORE: All CPUs Content

Rise of the Tomb Raider & The Witcher 3: Wild Hunt

Rise of the Tomb Raider

Rise of the Tomb Raider has long been a thorn in Ryzen's side due to architectural eccentricities. Recent game patches have cleared up most of the inexplicable anomalies, thankfully.

The tuned Threadripper 1950X in Game Mode provides solid performance in this title, outpacing the less complex 1800X.

Intel's Core i7-7700K continues to leverage a single-threaded performance advantage, though an overclocked -7900X offers similar performance.

The Witcher 3: Wild Hunt

That large frame time spike in the benchmark's early section is a scene transition. All of our contenders suffer to some extent during this switch, though AMD's Threadripper 1950X stumbles more than the others.

Regardless, the overall delta between processors in our test pool is relatively small, and the 1950X offers a competitive 99th percentile frame time measurement.

MORE: Best CPUs

MORE: All CPUs Content

DTP, Office, Multimedia & Compression Performance

Although we usually don’t run our application benchmarks on overclocked processors, we're including the Ryzen Threadripper 1950X at stock and overclocked frequencies this time, since we wanted to know how AMD’s architecture scales in different scenarios.

To be fair, we also added a (reasonably) overclocked Intel Core i9-7900X to our results. This makes for an interesting comparison, especially when we see later that both AMD’s and Intel’s flagship processors break the 250W barrier during a rendering workload.

DTP & Presentation

Adobe’s Creative Cloud gives us a look at single- and multi-core performance. As such, it beats synthetic benchmarks as a productivity test.

After Effects CC is a classic when it comes to parallelized tasks, with the number of cores easily being more important than frequency. Conversely, InDesign CC shows Intel’s Skylake-X in the back of the pack, whereas Core i7-7700K scores major points with its high clock rate. Likewise, AMD’s Ryzen 7 beats Ryzen Threadripper, regardless of frequency.

Adobe’s Illustrator CC leaves us scratching our heads in confusion: the Ryzen Threadripper 1950X lands in dead last. Overclocking doesn’t make much of a difference.

Encoding & Multimedia

It’s Threadripper’s time to shine in our HandBrake benchmark. AMD’s processors just crunch those numbers no matter the selected quality setting (even though Intel’s Core i9-7900X ends up right between the Threadripper CPUs with HandBrake set to normal quality).

The order changes at the more demanding high-quality setting. Intel’s large processor can only keep up if it’s overclocked quite a bit first.

Compression & Decompression

AMD's Threadripper processors continue their number-crunching winning streak when it comes to compressing large files. Again, they leave Intel’s Core i9-7900X in the dust, until it gets overclocked.

Intel’s Core i7-7700K easily beats the entire field in our decompression benchmark due to its higher frequency. Still, AMD’s Threadripper processors do well, and they once again best Intel’s Skylake-X flagship at stock clock rates.

MORE: Best CPUs

MORE: All CPUs Content

2D & 3D Workstation Performance

2D Workstation Performance

Tom’s Hardware Synthetic 2D Benchmarks

We take a look at direct device write throughput first. The graphics driver uses the CPU heavily for this task, but doesn’t employ many threads. There hasn’t been true 2D hardware acceleration since the introduction of the unified shader architecture, after all. Microsoft's Windows driver model provides a huge obstacle for 2D hardware acceleration as well.

We up the ante by introducing memory to the mix. This is done with the help of the only remaining 2D hardware function: generating the graphics output in memory and then copying it to the output device all at once. The benchmark’s the same as before. We just plot a bitmap in memory, as opposed to sending the information directly to the monitor. The bitmap’s copied to it only once it’s complete. This pushes the CPUs, since they’re no longer platform-bound.

The results are surprising: AMD’s Ryzen Threadripper dominates the field, including the Core i7-7700K that’s usually a strong contender.

AutoCAD 2016 (2D)

Even though AutoCAD does use DirectX, ultimately it just duplicates every single draw function in software. The results are exactly as expected, and IPC throughput is emphasized due to AutoCAD’s limited scaling with additional cores.

3D Workstation Performance

AutoCAD 2016 (3D)

Clock rate trumps core count. AMD’s Ryzen 7 and Ryzen Threadripper end up fairly close to each other. AutoCAD’s performance turns out to be close to that of older games, since it uses DirectX and isn’t really optimized to take advantage of multiple cores.

Cinebench R15 OpenGL

The Cinebench R15 OpenGL benchmark demonstrates what happens when software isn’t optimized for AMD’s Ryzen processors. Clock rate edges out core count, and Cinebench R15 OpenGL retains its tendency to favor Intel processors.

SolidWorks 2015

The same goes for SolidWorks 2015. Even a Threadripper 1950X processor overclocked to 3.8 GHz loses to a Ryzen 7 1800X. Changing the memory access setting in the BIOS would result in parity. However, doing so would also negatively impact most compute tasks. It’s one or the other, and we can’t have both. This is a hard decision that comes down to which of the two is more important.

Creo 3.0

The trend continues with Creo 3.0. For AMD’s Threadripper processor, core count edges out clock rate this time around. Still, it ends up at the bottom of this 3D benchmark heap. Optimized BIOS settings would push it up to the middle of the field, but would also negatively impact the CPU composite score, which can be found on the next page.

Blender (Real-time 3D Preview)

Notwithstanding Intel's Core i7-7700K, the Blender benchmark results are acceptable. Taken together with the fantastic rendering performance, which is shown on the following page, things look good.

Catia V6 R2012

This is one of the graphics benchmarks that has been optimized time and again (it’s part of the free SPECviewperf 12 suite). Clock frequency is important to this metric.

Maya 2013

The same can be said for Maya 2013. Note that the real-time 3D output numbers don’t tell the whole story. AMD’s Threadripper processors fare a lot better when it comes to final rendering.

For pure design and drafting work, AMD’s Ryzen Threadripper 1950X is a respectable option. It isn't great either, though. In the end, it can be used for these professional tasks if you simultaneously hit it with other workloads that need to be completed at the same time. Otherwise, fewer cores operating at higher clocks tend to offer better performance.

MORE: Best CPUs

MORE: All CPUs Content

CPU Computing & Rendering Performance

CPU Workstation Performance

Many modern suites include modules that are based exclusively on computing and simulations. This means we need to go beyond just 3D workstation performance to form our opinion of these high-end CPUs.

SolidWorks, for instance, doesn't scale well with increasing core count, which means that even quad-core processors with high IPC (and SMT) do well. AMD’s Ryzen Threadripper processor keeps up with respectable results.

Frequency is what matters in Creo 3.0, so long as your CPU offers at least eight threads.

Clock rate and core count matter in 3ds Max 2015. Intel’s Core i7-7700K performs surprisingly well due to its high frequency. It would fare a lot worse if we turned off Hyper-Threading.

The CPU composite score includes rendering, which has its own separate section right below. Consequently, AMD’s Ryzen Threadripper processor does really well.

CPU Performance: Photorealistic Rendering

Final rendering doesn’t require a CPU that's good at everything. Rather, this task wants efficiency and fast parallel computation.

Nobody beats AMD’s Ryzen Threadripper when it comes to rendering in 3ds Max 2015. Core count is much more important than clock rate, and performance scales beautifully with added on-die execution resources.

The console version of LuxRender confirms these results. The 1950X is in a league of its own.

Last but not least, we take a look at Blender. The usual workload (with a sample size of 200 pixels) confirms what we saw in the preceding benchmarks. The Threadripper 1950X finishes way ahead of the field.

The results obtained from SPECwpc’s Blender loop look very similar, even though this benchmark presents a somewhat different task consisting of more than just rendering.

With the rendering portion of the workload easing up, a stock Intel Core i9-7900X rejoins the party.

This trend gets stronger once multiple factors play a role in the benchmark loop, not just photorealistic rendering. It’s not exclusively up to core count anymore, but IPC’s important as well in this scenario.

Intel’s Core i9 CPUs offer acceptable performance for the semi-professional field. However, practically the same can be said for AMD’s Ryzen Threadripper processor. Depending on the task, it ranges from being able to keep up reasonably well to beating the competition hands-down. We’ve waited a long time to say that. The accolades are well deserved!

MORE: Best CPUs

MORE: All CPUs Content

Scientific & Engineering Computations, & HPC Performance

For these tests, we’re using the SPECwpc benchmark suite for workstations with its wide variety of tasks. It tests a number of very different mathematical computations optimized for parallelization. They typically make heavy use of available memory bandwidth and cache, plus expose issues with latency.

Rodinia

The pre-Euler3D CFD test (Computational Fluid Dynamics benchmark) runs very well, with AMD’s Ryzen Threadripper posting great results. It’s interesting to see that the overclocked processors don’t really improve performance all that much.

Convolution

In this benchmark, a mathematical operation is performed on two functions (convolution), which results in a third function. Performance scales similarly well with core count and clock rate.

CalculiX

This test is based on the finite element method for three-dimensional structural computations. The Ryzen Threadripper 1950X does edge out Intel’s Core i9-7900X flagship.

Poisson's Equation

Poisson's Equation is a second-order partial differential equation widely used in physics for boundary value problems.

Overclocking AMD’s Threadripper processors doesn’t yield much of a performance increase. Meanwhile, Intel’s Core i9-7900X dominates the field. The Threadripper chip posts better results than Intel’s remaining Core i9 CPUs, but it needs a lot more cores in order to keep up.

Sequential Reweighted Message Passing (SRMP)

These are algorithms for discrete energy minimization. None of AMD’s processors do well. However, the Threadripper CPU is able to compensate to some degree using core count, whereas the Ryzen 7 CPUs fall way behind.

Kirchhoff Migration

The earth’s subsurface structure can be determined via seismic processing. One of the four basic steps in this process is the Kirchhoff Migration, which is used to generate an image based on the available data using mathematical operations. This benchmark and its underlying computations turn out to be a great fit for AMD’s Ryzen Threadripper processor.

The Threadripper 1950X is a good choice for these tasks. The only exceptions are software packages that just aren't well-optimized for it, such as the SRMP test. Surely, developers will be more cognizant of AMD's hardware moving forward, now that the company has a competitive architecture designed to tantalize this market.

Consequently, negative outliers should become a less and less frequent, just as they have on the desktop with Ryzen 7, 5, and 3. Bottom line: the Ryzen Threadripper 1950X is a great choice when computing power and easily parallelized tasks are involved.

MORE: Best CPUs

MORE: All CPUs Content

Overclocking, Cooling & Temperature

The Right Cooling Solution

AMD doesn't use thermal paste between the Ryzen Threadripper processors’ dies and integrated heat spreader. Instead, it went with good old solder. This decision should prove critical during our overclocking efforts.

Most reviewers (us included) received all-in-one liquid coolers for their Threadripper CPUs. They're made by Thermaltake and include a flat 360mm radiator with three 120mm fans.

Fully assembled, and sitting next to the motherboard and RGB-lit memory sticks, the whole kit lights up like a Christmas tree. At least it was (almost) fully functional. The only exception was the original thermal paste on Asetek's pump. There was simply not enough of it. Even without overclocking, AMD’s new processor can hit 180W under heavy load, and its heat spreader is relatively large. This combination requires a different approach. Instead of the usual centered blob, we drew a thick line with the thermal paste. We then put the pump in place and gently rotated it back and forth while manually applying some pressure. Only then did we screw it in place.

Installing Your Own Water Cooler

We’ve already reported how AMD changed its interface for Socket TR4 (SP3r2). One of the most important alterations involves the screws. AMD went with M3.5 screws, an uncommon size with a fine thread. You won’t find these at your local hardware store. The difference can easily be spotted in the picture below, with our purchase on the left and an original AM4 screw on the right:

Simply shipping brackets of a different size won’t do the trick here. The acquisition of suitable screws is also necessary. The socket’s threads don’t go all the way through, which is to say that they aren’t open in the back. So, the screws also need to be of the correct length. We tried 20mm screws, and they turned out to be too long. Then we experimented with fitting spacers. A good length for the screws would have been 15mm, but this might vary depending on the bracket’s thickness.

The next question, which is even more important than the first one, is how much pressure to apply. We used a special torque wrench with very small steps and M3.5 screws with internal hexagonal threads, and started with a reasonable 0.1Nm. From approximately 0.25Nm, we couldn’t detect any further gains in cooling performance, so that's where we stopped. A sensible maximum pressure is approximately 0.35Nm.

The picture below shows our completely installed cooling solution. The CPU block’s an Alphacool XPX with a bracket for AMD’s Socket TR4 (SP3r2). The polyamide washers are used as spacers and take the place of springs. The top washers are made of steel to keep the head of the cylinder from boring into the much softer washer material during the CPU block’s installation.

Overclocking

Our Ryzen Threadripper 1950X sample overclocked to 3.9 GHz and 1.35V. However, the all-in-one water cooling solution in our kit couldn’t keep the system stable at that level. This was due to the processors’ power consumption rising to more than 250W during rendering.

And so we went back to our Chiller to achieve some better comparisons. This way, there’s one true constant to our measurements: a water temperature of approximately 20°C, which can be held constant, even topping more than 300W of waste heat.

For everyday use, a normal water-cooling solution will definitely suffice thanks to the soldered heat spreader, which makes AMD’s processor much less of a challenge than Intel's Core i9-7900X. Our approach simply allows us a bit more sophistication in our overclocking endeavors.

Using the Chiller, AMD’s Ryzen Threadripper 1950X achieved 4 GHz at 1.45V. Just don't expect to see those results from your own overclock. Even the Chiller started to fall behind the CPU's thermal output as temperatures crept too high for comfort. This is why the processor is overclocked to more reasonable levels for our benchmarks, with Threadripper running at 3.9 GHz.

Maximum Temperatures: Stock Clock Rate

AMD circulated a 27°C addition to the Tctl values, which is supposed to amount to the average core temperature. This sounds about right to us after taking a look at the temperature difference between Tctl and Tdie (the latter being the chip temperature). Between this and the fact that a huge cooler made it practically impossible to conduct our own heat spreader measurements, we're forgoing the delta measurements you saw in our Ryzen 3, 5, and 7 reviews.

We stick with the water cooling solution that AMD provided for our first results. As a bit of a power consumption spoiler, the motherboard limits the two CPUs between 179 and 180W. This upper boundary can’t even be exceeded for short periods of time using normal motherboard settings.

Here are the temperature curves:

The CPU temperature values reported by HWInfo64 through Asus' separate sensor loop are between 6°C and 12°C lower than the Tctl values, and they rise more slowly. The voltage converter temperatures of just under 60°C, achieved without any additional cooling, are great.

Maximum Temperatures: Overclocked

Increasing voltages to guarantee stable operation pushes the processor well beyond its sweet spot. Consequently, power consumption goes through the roof. Operating well beyond 300W poses a challenge for any cooling solution. That's why we're using the Chiller. We did try a normal water-cooling loop though, resulting in the Tctl and Tdie values going up by ~10°C to 15°C. This is well within an acceptable range.

The overclocked 1950X peaks at a hefty 320 to 325W. Using the Chiller, this level of power consumption is accompanied by Tctl values of 87°C. That’s actually not as severe as it seems once the offset and Tdie values are taken into account. A real temperature of approximately 60°C serves as a great demonstration of why solder is superior to thermal paste. Intel's Core i9-7900X could have had so much potential if the company hadn't taken the easy way out.

Voltage Converter Temperatures

We’ll conduct a separate test with different loads and X399 motherboards in the future. As for the Asus X399 ROG Zenith, its voltage converters generally stay under 100°C without any additional air cooling, even as AMD's CPU consumes well over 300W.

Asus set its throttling temperature threshold to 105°C. Even a bit of airflow helps, though. This is demonstrated quite nicely by the temperature curves for the overclocked configuration above. The fan that’s installed right above the I/O shield doesn’t really have any effect, unless you count its unnecessarily high noise level. Asus should have done without this gimmick.

MORE: Best CPUs

MORE: All CPUs Content

Power Consumption

We establish the package’s power consumption results by using a special sensor loop. This way, our values represent the exact amount of power that goes into the CPU and then reemerges in the form of power loss, which is to say waste heat dissipated by the cooling subsystem. We check our sensor readings using shunts and measuring overall power consumption directly at the EPS connector (current probe and direct voltage measurement).

The CPU values supplied by the voltage converters on the motherboard via HWinfo64 are sufficiently precise for our purposes, even though the measurement intervals are longer than those of our oscilloscope. We retested Intel’s Core i9-7900X after overclocking it to 4.5 GHz and added its results to the graphs to provide a fair comparison.

We need to note that AMD’s Threadripper CPUs use different partial voltages for the SOC and SMU rails at different clock rates. These partial voltages, which, again, depend on the frequency, do have an influence on the overall package’s power consumption results. AMD recommended that we use the profile included with their 32GB DDR4-3200 kit. If the memory was operated using the standard SPD values for DDR4-2133, then the power consumption would be 15W lower!Both of AMD’s CPUs are designed for a maximum power consumption of 180W at their default settings. If the memory’s overclocked, then the CPU gets 15W less, which might affect performance in usage scenarios that employ all of the cores and, consequently, get too close to the limit.

Idle Power Consumption

Compared to AMD’s Ryzen 7, the Threadripper’s idle power consumption is just over twice as high. However, Threadripper also hosts two dies instead of one, and it also hits higher clock rates under occasional loads. The overclocked version utilizes higher voltages as well, and we just mentioned the memory’s role in power consumption.

CAD Workload Power Consumption

AutoCAD 2016 rarely uses more than two or three cores for its usual tasks. In fact, most of the time it's limited to a single core. Thus, it's not surprising that the CAD power consumption only adds a maximum of 15W to the idle power numbers, especially since the latter’s not truly idle power consumption to begin with.

The two overclocked versions add another 14W, which makes for an almost 30W difference compared to the idle power consumption. And then there’s the memory, of course.

Gaming Power Consumption

When it comes to gaming, we see an issue where Threadripper’s many cores get in each others' way. Consequently, overall performance ends up lackluster. The power consumption’s on the same level as that of Intel’s Core i9-7900X, even though Skylake-X fares much better in the benchmarks. The results are similar to those we reported for the real-time 3D CAD benchmarks.

Stress Test & Maximum Power Consumption

Power consumption goes through the roof during our stress test. This is especially true for the overclocked configurations.

In the case of a stock Intel Core i9-7900X, the motherboard has to shoulder some of the blame for this. It doesn’t lower the processor’s clock rate in accordance with the rules, but leaves them at a much higher level.

AMD’s Ryzen Threadripper doesn’t have those kinds of issues. The Asus X399 ROG Zenith Extreme motherboard limits power consumption to exactly 180W, just as it should, when using the default settings. Things look a whole lot different once the processor is manually overclocked to maximize its frequency, though. The 1950X needs 1.35V to achieve 3.9 GHz. At that point, AMD's new processors join Intel's overclocked Core i9-7900X well beyond 300W.

Bottom Line

Depending on the task, Threadripper's two dies sometimes consume more power than other processors’ single dies. This is to be expected when all cores are operating at full load. The high power consumption’s okay if it’s put in the context of correspondingly high application performance, as long as the work that’s being done has been parallelized well and the software is optimized for AMD's architecture.

Unfortunately, Threadripper's efficiency during gaming turns out to be significantly worse than Intel’s. Skylake enjoys notably higher IPC, after all. A large portion of Threadripper's higher power consumption comes simply from a baseline that notches up from Ryzen 7. Even at idle, it draws an additional 15W or so. If those 15W are subtracted from Threadripper's gaming power consumption, then the results look a lot better. It brings their lower power consumption increase compared to idle in line with their lower gaming performance.

MORE: Best CPUs

MORE: All CPUs Content

Final Analysis

AMD’s mainstream Ryzen CPUs heralded the return of competition in the desktop processor market. Now, AMD brings the same Zen architecture and strategy to the high-end desktop, and we’ve already seen Intel’s reaction in the form of lower prices for its (still-pricey) Skylake-X line-up.

First, the elephant in the room: AMD positions Threadripper for creators, heavy multi-taskers, and gamers who stream to services like Twitch. It also specifically states that the processor isn’t intended for low-resolution gaming, particularly under lightly-threaded titles.

We are going to explore more intense use-cases in an upcoming feature, but were unable to complete streaming testing due to the usual time constraints we face during a launch window. We did run some ad hoc tests and were able to easily play Mafia 3 at 4K while encoding a video and running a virtual machine, which we could still access via remote desktop. We didn’t experience any significant performance degradation via our own subjective measure. Still, we prefer hard data and will work at putting definitive results behind the experience.

High-resolution gaming is unequivocally a niche, according to Steam's Hardware and Software Survey, but it's clearly growing in popularity. Of course, enthusiasts who spend $1000 on a processor are far more likely to use high-resolution monitors. But testing at 1440p or 4K very obviously pushes the bottleneck back over to the GPU. And because we're testing CPUs in these reviews, we deliberately use 1080p as a tool to emphasize broad differences between architectures and more specific deltas separating models. As GPUs evolve, higher resolutions should become more processor-bound, and hopefully by then we'll see more optimization for lots of cores.

It’s clear that Intel's Core i9-7900X offers better average frame rates during purely gaming workloads, but our standard practice of focusing on 99^th percentile metrics takes performance and smoothness into consideration. We’re looking at a five FPS delta between the stock -7900X and Threadripper’s best stock configuration in all games, and six FPS for new games. That gap becomes seven and four FPS, respectively, after overclocking both processors. You can imagine that gap will shrink at higher resolutions.

We didn’t add platform costs to our price efficiency charts because all high-end parts drop into obviously premium platforms, and X399 is no exception. But be mindful that you'll pay a lot more for a HEDT platform than the two mainstream configurations we tested. Cheaper alternatives like the Core i7-7700K and Ryzen 7 series are likely better for the folks who are interested in gaming, first and foremost. Much like our recommendations for Intel’s high-end desktop processors, we don’t recommend AMD’s flagship 1950X for strictly gaming, either.

If your workloads are CPU-bound, though, Threadripper shines in our benchmarks written to exploit as many cores as you can throw at them. Threadripper outpaces the similarly-priced -7900X in rendering, encoding, and compression. As expected, it isn’t quite as nimble in lightly-threaded applications, such as decompression and portions of the Adobe suite. Those applications continue to favor Intel’s IPC throughput and frequency.

After the Ryzen launch, AMD was faced with the challenge of quickly maturing its motherboard ecosystem and convincing game developers to optimize existing titles for the new architecture. The company has met with success on many fronts in a relatively brief time (it’s hard to believe it’s only been five months), and enablement continues. Threadripper is a unique product that introduces even more complex challenges. No doubt, AMD is ready to take action on those, too. Case in point: Threadripper offers so many cores that some games won't even load. No doubt, Intel will face the same conundrum in the future as it scales out its architectures as well.

AMD is obviously aware of the challenges it faces. Using a combination of BIOS switches and Windows-based utilities, it exposes several knobs that ensure compatibility and address the architectural eccentricities of a data center-inspired desktop product. We’re sure to see well-heeled enthusiasts work through the settings to find the best combinations, even if most want to use Threadripper the way it ships. Of course, we like to experiment, so we’ll spend the coming weeks working on more stressful use-cases and finding the best combinations for different workloads.

Ryzen Threadripper 1950X is a solid entrant for AMD, and the company knows it's going after a niche market here. Those who need what Threadripper offers likely already know. And if that's you, we have to imagine you're elated to know there's an alternative to Intel's steep buy-in, particularly now that AMD is winning in benchmarks it hasn't won in a very long time.

MORE: Best CPUs

MORE: All CPUs Content

AMD Threadripper: Unboxing

Igor Wallossek — Thu, 03 Aug 2017 13:00:00 +0000

If the postman rings three times...

Yes, we were looking out the window, awaiting the Ryzen Threadripper's arrival. Finally! This time around, after only a one-day delay, we received two large packages with interesting contents.

MORE: Best CPUs

MORE: Threadripper Lands August 10, AMD Unveils Pricing, Accessory Kit, New 8-Core Model

MORE: Exclusive Threadripper Socket TR4 Schematics, Cooler Compatibility

Unforgettable Comes Often

But why two packages? We really only expect...

MORE: Best CPUs

MORE: Threadripper Lands August 10, AMD Unveils Pricing, Accessory Kit, New 8-Core Model

MORE: Exclusive Threadripper Socket TR4 Schematics, Cooler Compatibility

The Pelican Case Lands

... two CPUs, one motherboard, and maybe some memory. Inside the first package, we found a nice Pelican case with the Tom's Hardware logo emblazoned on the metal faceplate, along with AMD's Ryzen Threadripper branding.

MORE: Best CPUs

MORE: Threadripper Lands August 10, AMD Unveils Pricing, Accessory Kit, New 8-Core Model

MORE: Exclusive Threadripper Socket TR4 Schematics, Cooler Compatibility

The Pelican's Innards

A few snaps of the latches later, we found two big glowing eyes peering at us questioningly from the inside of the case. We couldn't quite capture the lighting effect with the camera, but when you open the case, and the lights come on, you know it's showtime! The extra crystal bling at the bottom of the case comes with a third Ryzen Threadripper CPU.

MORE: Best CPUs

MORE: Threadripper Lands August 10, AMD Unveils Pricing, Accessory Kit, New 8-Core Model

MORE: Exclusive Threadripper Socket TR4 Schematics, Cooler Compatibility

Lights, Camera, Action!

Without LED, there is nothing. We just had to find the light source, so we tore the case apart. AMD created the glowing eye effect with a few nice, bright rings that shine through the front of the two Threadripper packages. The switch up top triggers the effect, obviously.

MORE: Best CPUs

MORE: Threadripper Lands August 10, AMD Unveils Pricing, Accessory Kit, New 8-Core Model

MORE: Exclusive Threadripper Socket TR4 Schematics, Cooler Compatibility

The Enlightenment!

AMD thoughtfully silk-screened the display processor with the Tom's Hardware logo, but we still aren't sure if it's working silicon. Would you risk one of the few X399 motherboards on the planet (and our only one) just to find out? We're still pondering it.

It really is a work of art; it almost looks like a diamond in our favorite jeweler's shop window.

MORE: Best CPUs

MORE: Threadripper Lands August 10, AMD Unveils Pricing, Accessory Kit, New 8-Core Model

MORE: Exclusive Threadripper Socket TR4 Schematics, Cooler Compatibility

All Good Things Come In Twos

AMD sends samples to an incredible number of reviewers, so it's best to provide an identical test system to everyone. That way, the components are verified to be compatible, which eases troubleshooting if something goes wrong during testing. We also found a nice plexiglass poster of sorts marked with AMD's Ryzen messaging at the top and the requisite legal mumbo jumbo at the bottom.

Our Threadrippers came with an Asus ROG X399 Zenith Extreme motherboard, matching 16GB G.Skill DDR4-3200 (14-14-14-36) memory kits, a Thermaltake AiO 360 water cooler, and a 1,250W Thermaltake power supply. The remaining cartons inside are empty—but we checked just to make sure.

MORE: Best CPUs

MORE: Threadripper Lands August 10, AMD Unveils Pricing, Accessory Kit, New 8-Core Model

MORE: Exclusive Threadripper Socket TR4 Schematics, Cooler Compatibility

The Complete Package

RGB is more fun then ever before. Except for the CPUs themselves (why not, really?), all of the components are fully illuminated. We've got the sunglasses (welder's helmet?) ready for the tests, but we're pretty sure you'll be able to see our test bench from space with the naked eye.

MORE: Best CPUs

MORE: Threadripper Lands August 10, AMD Unveils Pricing, Accessory Kit, New 8-Core Model

MORE: Exclusive Threadripper Socket TR4 Schematics, Cooler Compatibility

Together In The Lab

We'll test the Ryzen Threadripper 1950X (16 cores / 32 threads) and the 1920X (12 cores / 24 threads), and perhaps the Tom's Hardware processor, in the upcoming review. There is no doubt that AMD has dropped a packaging bomb on Intel. The packaging is amazing—it is the most expensive we've seen in years.

These aren't special Tom's Hardware-edition boxes, either; these are the retail packages.

MORE: Best CPUs

MORE: Threadripper Lands August 10, AMD Unveils Pricing, Accessory Kit, New 8-Core Model

MORE: Exclusive Threadripper Socket TR4 Schematics, Cooler Compatibility

RIP the Threadripper

"Rip here," you say? We get it. Threadripper. It really is a shame to defile such a nice package, though, so we ripped with care. We wonder if you'll need that sticker for future mail-in rebates, though.

MORE: Best CPUs

MORE: Threadripper Lands August 10, AMD Unveils Pricing, Accessory Kit, New 8-Core Model

MORE: Exclusive Threadripper Socket TR4 Schematics, Cooler Compatibility

Unlocking The Goodness

No, that's not a ring of cheese, but a big plastic screw that binds the Styrofoam shell together. A quick twist...

MORE: Best CPUs

MORE: Threadripper Lands August 10, AMD Unveils Pricing, Accessory Kit, New 8-Core Model

MORE: Exclusive Threadripper Socket TR4 Schematics, Cooler Compatibility

Disassembled And Excited

...and even after the case splits open, we still have not reached the object of our desire. But we have learned one thing: Threadripper is like a shellfish that refuses to come out of its shell.

There isn't an AiO cooler as one of the recent rumors claimed, but there are a few other goodies. The manual sits atop the Threadripper-branded Torx driver you'll need for the socket. It also acts as a torque wrench to ensure you don't over-tighten the socket fasteners. There'sa watercooling adapter nestled in the Styrofoam, as well.

MORE: Best CPUs

MORE: Threadripper Lands August 10, AMD Unveils Pricing, Accessory Kit, New 8-Core Model

MORE: Exclusive Threadripper Socket TR4 Schematics, Cooler Compatibility

It Takes Some Work...

After another turn to the left, we finally reach the nectar-sweet center. We can finally see Threadripper in the flesh (er, silicon) after the orange insert pops free from the rear of the inner package. AMD gets an A+ for presentation.

MORE: Best CPUs

MORE: Threadripper Lands August 10, AMD Unveils Pricing, Accessory Kit, New 8-Core Model

MORE: Exclusive Threadripper Socket TR4 Schematics, Cooler Compatibility

The Eye Of Sauron

Ryzen's home looks like an old-school CRT television displaying the Eye of Sauron. The iris gazes down on the stickers and documents that came in the box. We don't mention packaging in CPU reviews because it's usually so boring, but AMD has changed the game. Every item is well thought out and designed to the point of being a collector's item. Of course, AMD hopes to sell millions of these packages, so resale value won't be high. It'll sure look good on a mantel or shelf, though.

MORE: Best CPUs

MORE: Threadripper Lands August 10, AMD Unveils Pricing, Accessory Kit, New 8-Core Model

MORE: Exclusive Threadripper Socket TR4 Schematics, Cooler Compatibility

The Hockey Puck

We popped off the metal clamp, which took some doing, then removed the black outer casing to finally reveal the Threadripper 1950X sitting snugly in its hockey-puck-like casing. A tug on the tab lifts the processor up, and then you slide it out of the case at an angle.

MORE: Best CPUs

MORE: Threadripper Lands August 10, AMD Unveils Pricing, Accessory Kit, New 8-Core Model

MORE: Exclusive Threadripper Socket TR4 Schematics, Cooler Compatibility

And There It Is

The carrier frame has a nice little orange arrow in the corner, just like most common processors. That'll help make sure you insert the processor correctly into the carriage tray, but it's pretty hard to mess up with the new design. The frame stays attached to the processor when you slip it into the socket, so it must be somewhat heat resistant. But how much heat does Ryzen Threadripper generate? We'll get to that in the review.

For most, the bright orange carrier frame and processor will head straight to the motherboard's socket, but we've got a few other ideas.

MORE: Best CPUs

MORE: Threadripper Lands August 10, AMD Unveils Pricing, Accessory Kit, New 8-Core Model

MORE: Exclusive Threadripper Socket TR4 Schematics, Cooler Compatibility

Who Is The Greatest?

AMD's Ryzen Threadripper wins the "battle of the biggest" against Intel's Skylake-X, but who is the real heavyweight?

MORE: Best CPUs

MORE: Threadripper Lands August 10, AMD Unveils Pricing, Accessory Kit, New 8-Core Model

MORE: Exclusive Threadripper Socket TR4 Schematics, Cooler Compatibility

The Weigh In

We stripped the processors down and threw them on the scale (but not soaking wet). In the blue corner, we find the Intel Core i9-7900X, the summer heavyweight champion. With a total of 61.4 grams (~2.2 oz.), he seems to have missed his weight class by a little bit. Perhaps the heat has gotten to him.

MORE: Best CPUs

MORE: Threadripper Lands August 10, AMD Unveils Pricing, Accessory Kit, New 8-Core Model

MORE: Exclusive Threadripper Socket TR4 Schematics, Cooler Compatibility

Heavyweight In The Red Corner

The 134.9 gram (~4.76 oz.) giant from the red corner is more than twice as heavy as the reigning high-end desktop champion. This raises a few questions. Heavier, in this case, doesn't mean more expensive, but does it mean faster? We're certainly going to find out.

MORE: Best CPUs

MORE: Threadripper Lands August 10, AMD Unveils Pricing, Accessory Kit, New 8-Core Model

MORE: Exclusive Threadripper Socket TR4 Schematics, Cooler Compatibility

Full Frontal

The Threadripper 1950X's dimensions are almost gigantic (it has to be to fit the word "Threadripper" on there, after all).

MORE: Best CPUs

MORE: Threadripper Lands August 10, AMD Unveils Pricing, Accessory Kit, New 8-Core Model

MORE: Exclusive Threadripper Socket TR4 Schematics, Cooler Compatibility

LGA To The Extreme

AMD's Ryzen Threadripper snaps into the 4,094-pin TR4 LGA (Land Grid Array) socket. Skylake-X's socket is puny in comparison, with just 2,066 pins, which is a little more than half the size of AMD's TR4 behemoth.

MORE: Best CPUs

MORE: Threadripper Lands August 10, AMD Unveils Pricing, Accessory Kit, New 8-Core Model

MORE: Exclusive Threadripper Socket TR4 Schematics, Cooler Compatibility

The Square Things

We see that the processor's heat spreader is significantly larger than a normal CPU water block. But for our first tests, we'll feed the Alphacool XPX waterblock with a chiller.

MORE: Best CPUs

MORE: Threadripper Lands August 10, AMD Unveils Pricing, Accessory Kit, New 8-Core Model

MORE: Exclusive Threadripper Socket TR4 Schematics, Cooler Compatibility

The Round Things

AMD includes a bracket for round Asetek pumps, and you can also use it with many AIO compact water coolers. The AIO's usable area on the heatspreader is even smaller, but we're optimistic that it will work (for the time being). After all, AMD has decided to provide us with a 360 radiator, which should help to maximize the small cooling footprint. We've already posted exclusive schematics of the TR4 socket, which outlines some of the challenges associated with cooling the massive processors.

There are also a number of new air coolers on the market (Noctua's got a few beastly models), but it's only a matter of time until we see special "Threadripper Edition" TR4 watercoolers that make contact with the entire IHS.

MORE: Best CPUs

MORE: Threadripper Lands August 10, AMD Unveils Pricing, Accessory Kit, New 8-Core Model

MORE: Exclusive Threadripper Socket TR4 Schematics, Cooler Compatibility

Hands-On Testing

Now that we know the official combat weight, we provide the promised hands-on test. As you can see, our hand is on it.

In time, we will cover the rest in detail, as usual. We promise.

MORE: Best CPUs

MORE: Threadripper Lands August 10, AMD Unveils Pricing, Accessory Kit, New 8-Core Model

MORE: Exclusive Threadripper Socket TR4 Schematics, Cooler Compatibility

Into The Socket

You won't find Intel's expensive latches on this X399 motherboard. Instead, Socket TR4 has three Torx screws and two frames. The new mounting mechanism is much more refined than previous designs; you simply slot the CPU right into a carrier frame and tighten it down. Head over to MSI's video of the Threadripper installation procedure in the TR4 socket for more detail.

AMD also hid a 512GB M.2 Samsung Pro SSD under the motherboard's heatsink. That's yet another nice touch—provided you know it's there. If you don't, you will be greeted with unexplained HDD error messages on the motherboard's LED screen; these are seemingly not correctable. That might trigger a bit of anxiety until you figure it out.

Not that we would know.

MORE: Best CPUs

MORE: Threadripper Lands August 10, AMD Unveils Pricing, Accessory Kit, New 8-Core Model

MORE: Exclusive Threadripper Socket TR4 Schematics, Cooler Compatibility

Extra-Large-Grid-Array

If you like, you can count how many pins you see here. If you're in a hurry, though, it is 4,094...

MORE: Best CPUs

MORE: Threadripper Lands August 10, AMD Unveils Pricing, Accessory Kit, New 8-Core Model

MORE: Exclusive Threadripper Socket TR4 Schematics, Cooler Compatibility

Now That's Tight

The processor is locked in, tightened down, and bomb-proof. There shouldn't be any contact problems.

MORE: Best CPUs

MORE: Threadripper Lands August 10, AMD Unveils Pricing, Accessory Kit, New 8-Core Model

MORE: Exclusive Threadripper Socket TR4 Schematics, Cooler Compatibility

Finished, For Now

And that's it for now. If you've got an appetite for more, we'll reveal the purpose behind this unboxing soon enough, so please be patient. August 10 is only seven days away...

MORE: Best CPUs

MORE: Threadripper Lands August 10, AMD Unveils Pricing, Accessory Kit, New 8-Core Model

MORE: Exclusive Threadripper Socket TR4 Schematics, Cooler Compatibility

Threadripper Lands August 10, AMD Unveils Pricing, Accessory Kit, New 8-Core Model

palcorn@outlook.com (Paul Alcorn) — Mon, 31 Jul 2017 02:30:00 +0000

Threadripper lands on retail shelves on August 10 and pre-orders open tomorrow. AMD filled us in on additional details at its Capsaicin event here in Los Angeles today, including an unannounced processor, a new accessory kit, and Threadripper's XFR frequencies and TDPs.

AMD's Zen microarchitecture, which was first officially available in early March, has truly revitalized the company's processor lineup. The design can scale from low-power mobile models to the high-powered data center, so it was only a matter of time before AMD expanded its attack into Intel's high end desktop stronghold.

AMD's halo Ryzen Threadripper 1950X model leads the charge with 16 cores and 32 threads, and the Zen architecture bristles with copious PCIe connectivity options for the entire lineup. More importantly, Threadripper sets the stage for a potentially lopsided fight between this $999 16-core flagship model and the similarly priced Intel 10-core i9-7900X.

Lower prices and less segmentation are music to enthusiast ears, but we still have a lot of ground to explore when it comes to performance. Here's what we know now.

The Lineup

AMD aims the Threadripper lineup at software developers, video/audio engineers, and of course gamers, particularly those who stream or multitask heavily. The competitive Ryzen 7 lineup addressed the lower spectrum of professional applications, but dual-channel memory limited its reach. Threadripper's step up to a quad-channel memory controller provides enough memory heft to handle voluminous professional applications. Throwing in a beefy slab of PCIe connectivity also provides plenty of NVMe storage ports and supports up to four GPUs for heterogeneous computing.

The $999 16C/32T Threadripper 1950X is disruptive, but AMD also offers two competitive downstream models. The $799 1920X wields 12C/24T with a slightly higher base clock, and due to the basic rules of semiconductor power and thermal scaling, the $549 8C/16T 1900X features the highest base clock of the family. AMD already spouted the basic specifications of its two high-end models, but the 1900X was somewhat of a surprise announcement.

	Threadripper 1950X	Core i9-7900X	Threadripper 1920X	Core i7-7820X	Threadripper 1900X
Price	$999	$999	$799	$599	$549
Interface/Chiset	TR4 / X399	LGA2066 / X299	TR4 / X399	LGA2066 / X299	TR4 / X399
Cores/Threads	16/32	10/20	12/24	8/16	8/16
TDP	180W	140W	180W	140W	180W
Base Frequency (GHz)	3.4	3.3	3.5	3.6	3.8
Boost Frequency (GHz)	4.0 (4.2 XFR)	4.3 / 4.5 (TB 3.0)	4.0 (4.2 XFR)	4.3 / 4.5 (TB 3.0)	4.0 (4.2 XFR)
L3 cache (L2+L3)	40 MB	23.75 MB	38 MB	19MB	20 MB
Memory Support	DDR4-2667	DDR4-2666	DDR4-2667	DDR4-2666	DDR4-2667
Memory Controller	Quad Channel	Quad Channel	Quad Channel	Quad Channel	Quad Channel
Unlocked Multiplier	Yes	Yes	Yes	Yes	Yes
PCIe Lanes	64	44	64	28	64

The Threadripper 1950X naturally competes with Intel's Core i9-7900X, but it brings more cores, cache, and PCIe lanes to the battle. The Threadripper 1920X straddles the pricing line between the Core i9-7900X and the i7-7820X, which shows that AMD is taking advantage of the big $400 price gap in the Intel lineup. Shrewd move; Intel doesn't have a clear contender at this price point.

The Threadripper 1900X only undercuts the -7820X by $50, but it arrives at the party with 64 PCIe lanes. As such, it poses a real threat to Intel's Core i7-7820X.

AMD also announced that it offers a 200 MHz XFR boost, which brings the top stock frequencies to 4.2 GHz for all Threadripper processors. That narrows the gap between AMD and Intel's boost frequencies, but AMD also offers higher base frequencies across the stack.

AMD also provided a slide that outlines Threadripper's resource advantages relative to Intel's lineup. Overall, AMD's core, cache, and PCIe advantages are impressive. Although these factors don't always equate to better performance, they surely set a strong foundation. AMD also reminds us that it dedicates four of Threadripper's 64 PCIe lanes to the chipset.

Unleashing The Threads

AMD also treated us to performance comparisons, but as with all vendors, we have to take the claims with a grain of salt. We provide expandable test notes at the bottom of the article.

Intel's Skylake-X models have leading IPC throughput and boost clocks, but it isn't enough to hold off AMD's 1950X in multithreaded workloads. The 1950X's 60% core count advantage yields a 39% Cinebench performance advantage. The 1920X, which is 20% cheaper than the -7900X, provides 11% more multithreaded performance. AMD provided performance data from a broader spate of heavy workloads for the 1950X and 1920X, and even the 12-core 1920X can trade blows with Intel's $999 Core i9-7900X, according to AMD's numbers. We will have to wait for more information on the eight-core 1900X.

AMD also claims a performance-per-Watt advantage over comparably-priced Intel models. We learned the Threadripper models all have a 180W TDP, which is 40W more than the Skylake-X lineup, but TDP figures can be misleading. For instance, AMD's EPYC processors have a higher TDP than Intel's Purley models but are surprisingly efficient. Those processors are very similar to Threadripper.

AMD contends that its 16-core 1950X offers considerably more efficiency than Intel's 10-core -7900X, but it will also be interesting to pit it against one of Intel's higher core count models. Those have yet to come to market, lending AMD the advantage of the highest core counts at launch.

The Launch Ecosystem And Tool Kit

AMD is eager to avoid some of the hiccups of the Ryzen 7 launch, which suffered from limited motherboard and cooling options. To that end, AMD has lined up launch-ready motherboards from ASRock, Gigabyte, Asus, and MSI.

A menu of 20 different liquid cooling options and five air coolers from a wide array of manufacturers also greet early adopters. AMD includes a bundled Asetek adapter for existing AIO watercoolers. Considering Asetek's role in the AIO ecosystem, the bracket will grant wide compatibility with existing closed loop coolers.

Threadripper's large retail package provides plenty of room for value-adds. Given the large Threadripper processors and TR4 socket, it will likely require a fair amount of pressure to ensure consistent contact between the chip and the socket, so the company employs three Torx fasteners to hold the processor snugly. AMD provides a bundled Torx wrench for processor installation. We've covered the socket dimensions and installation procedure, so head to those articles for more detail.

Pre-orders and Availability

AMD is opening a pre-order window for the Threadripper 1950X and 1920X models on July 31, but as usual, we recommend waiting for in-depth reviews before pulling the trigger. For those who can't wait, Amazon, Newegg, and a host of other retailers will have the processors. A slew of boutique builders already offer pre-orders for full systems. The 1950X and 1920X ship on August 10, and the eight-core Ryzen Threadripper 1900X will land on August 31.

AMD Ryzen Threadripper 1950XView Deal

AMD Ryzen Threadripper 1920XView Deal

The Intel Reaction

AMD feels that Intel has already begun to lower its prices in the face of stiffer competition, and as we can see in the chart, Intel has significantly reduced its pricing for high-end desktop models. If AMD generates enough sales, it's possible we could see further changes to Intel's pricing model.

It's definitely been an exciting (and exhausting) year of processor releases so far, but the story isn't over. AMD also has the Ryzen Mobile and Raven Ridge APUs, which feature Zen and Vega cores on the same die, headed to market later this year. That puts a cap on AMD's penetration into every market segment. AMD has built its way up the stack to Threadripper, and you can imagine reviews will be ready when the processors hit shelves.

MSI Video Outlines Threadripper Installation Procedure In X399's TR4 Socket

palcorn@outlook.com (Paul Alcorn) — Fri, 28 Jul 2017 22:52:01 +0000

MSI released a video outlining the mounting procedure for AMD's 16-core/32-thread Threadripper processors that find a home in AMD's new TR4 socket on X399 motherboards. Cramming 16 cores into a single processor is quite the feat, but it also creates a very large chip. That means AMD had to design a massive new socket for Threadripper, but ensuring a snug fit for reliable communication between the processor and the motherboard presents engineering challenges. We recently posted exclusive schematics of the socket and how it will work with existing cooling solutions, but MSI's video brings a few more details to light.

AMD's new Socket TR4 (SP3r2) for consumer motherboards boasts 4,094 pins, which is why it's so much larger than Intel's 2,066-pin LGA2066 socket found on X299 chipset-equipped motherboards. That's a heaping ration of gold pins that result in a comparatively more complex design than Intel's socket. In fact, AMD's socket spans nearly the entire length of the DIMM slots that flank it on either side. Those factors might result in fairly expensive X399 motherboards, but we'll dive into that in more depth in the pending Threadripper review.

The TR4 socket marks AMD's transition from the PGA (Pin Grid Array) socket found on its AM4 motherboards to an LGA (Land Grid Array) interface. There are advantages and disadvantages to both designs, which again, we'll dive into in the review.

AMD's TR4 socket uses three Torx screws to keep Threadripper snug in its new home, which stands in contrast to the normal latching feature found on consumer motherboards from both Intel and AMD. In fact, it's similar to the technique Intel uses on its enterprise Purley processors that leverage the LGA 3467 Socket P, which features a comparably-beastly 3,647 pins. In contrast, Intel's mounting scheme employs Torx fasteners that mount through the heatsink and into the socket housing, whereas AMD's mounting scheme employs Torx screws that slot in directly through the flip-up socket housing. Head over to our Purley review for a look at the Intel mounting procedure, which requires you first to snap the processor into the heatsink then install the assembly onto the socket.

For Threadripper, first, you remove the three screws and swing open the retention mechanism, under which there is another assembly that you also swing open. You remove an internal plastic cover and then slide the processor—which is contained in an orange plastic shroud—into the carriage. It's important to leave the black cover over the socket until you install the processor in the flip-up housing—those pins are fragile.

After you slide the processor into its housing, you remove the cover and swing it down into the socket. Then you swing down the mounting mechanism and tighten the fasteners, after which you are ready to slather on a generous helping of TIM (necessitated by the large integrated heat spreader). Heatsink installation requires four additional fasteners. As we noted in our schematic piece, the heatsink mounting fasteners aren't aligned symmetrically.

Plug in the fan, and you're ready to rock. Threadripper is almost among us. AMD tells us that shipping begins in early August. As you can imagine, our preparatory testing has already begun.

AMD Ryzen Threadripper 1950XView Deal

AMD Ryzen Threadripper 1920XView Deal

Update, 7/28/17, 6:40pm PT: Corrected AMD core count

Exclusive Threadripper Socket TR4 Schematics, Cooler Compatibility

Igor Wallossek — Thu, 27 Jul 2017 23:38:02 +0000

AMD's 4,094-pin socket, which is paired with the X399 chipset for Threadripper, is massive by any measure. AMD christened it as Socket TR4 (SP3r2) for consumer motherboards and Socket SP3 for EPYC servers. The physical dimensions of both LGA (Land Grid Array) sockets are identical.

In either case, Socket TR4 far outweighs Intel's 2,066-pin socket for the X-Series platform, which has the natural side effect of a much larger heatspreader on the Threadripper processors. That presents challenges for cooler vendors, but AMD also uses indium solder under the heatspreader, which should help thermal performance significantly.

We know that AMD and various cooler manufacturers will be offering special solutions that will allow existing coolers to be used on AMD's much larger Socket TR4. We succeeded in obtaining a set of new adapters for the Alphacool XPX that we use in our German test lab. Please note that the manufacturer did not serve as a source for our exclusive schematics.

Since we now have the complete design drawings for the socket, we have superimposed the cooling solution over the socket schematic and made some educated guesses. This helps us to assess compatibility beforehand.

It's striking that, in addition to the rectangular shape of the CPU recess in the socket frame, there is an asymmetrical arrangement of the mounting screw emplacements. On the left, we see a distance of 65.2mm and to the right only 46mm. We also see very nicely where the modified square (or round heatsink) cooler will mount. A large, rather square waterblock, like the illustrated Alphacool XPX, should be an advantage over the round solutions we find on many of the Asetek products.

The rectangular heatsink covers about 90% of Threadripper's integrated heat spreader after accounting for the screws on the waterblock and the uncovered edges of the processor. Also, the arrangement and size of the microchannels inside the waterblock are, of course, not (yet) optimal. Threadripper CPUs have four die, and thus there are large disparate hotspots. This requires large-area microchannel fields and other flow profiles in the cooler.

For the launch, however, the water block should be sufficient, though it isn't perfect yet. We are curious when (and which) products will be specially adapted for the TR4/SP3r2 socket, and how much better they will perform.

For those of you do-it-yourselfers with a CNC milling machine, or for those who are simply curious, we have compiled all of the relevant design documents and present them below.

AMD Ryzen Threadripper 1950XView Deal

AMD Ryzen Threadripper 1920XView Deal

Socket SP3 front side

AMD Announces Ryzen Threadripper, Challenges Intel With 12C/16C For $799 and $999

palcorn@outlook.com (Paul Alcorn) — Thu, 13 Jul 2017 18:50:00 +0000

Enthusiasts have yearned for a suitable high-end desktop platform alternative for ages, and now we finally have one. AMD announced its highly anticipated Ryzen Threadripper in two flavors: The 1950X weighs in with 16 cores and 32 threads for $999, and the Threadripper 1920X brings 12 cores and 24 threads for $799. Both will ship in early August. AMD also announced its Ryzen 3 models and a ship date of July 27, but hasn't provided pricing information about these low-end models.

AMD released a video that outlines the basics of the launch and pits both Threadripper models against Intel's Core i9-7900X in a Multi-Threaded Cinebench R15 workload. The Ryzen Threadripper models win handily. That's a nice result, and it's noteworthy that both AMD models have the core count advantage, but the processors fall in the same price range. Given that Intel's 12-core and 16-core models aren't available yet, this is the closest we can get to a fair comparison to Intel's high-end models.

AMD's strategy has been to offer disruptive price points compared to competing Intel chips, and Threadripper's pricing continues that trend. Intel's competing Core i9-7960X features a similar 16C/32T as the $999 Threadripper 1950X, but it retails for $1,699. The $1,199 Core i9-7920X contends with the $799 12C/24T Threadripper 1920X.

Intel's realigned pricing for the Skylake-X processors was quite the buzz, but while the company has obviously reworked its high-end desktop pricing structure, it wasn't enough to stave off the underdog entirely. AMD's comparable products undercut Intel's Skylake-X models by $700 and $400, respectively, which is definitely going to fuel increased competition in the high-end desktop market, which has long been Intel's self-described high-margin lineup.

AMD released precious few details outside of pricing and frequencies. The Threadripper 1950X has a base 3.4GHz frequency that boosts to 4.0GHz under load. Current Ryzen models feature a two-core boost, and we expect that will continue. However, we have no information on XFR frequencies, which is an extra boost if your thermal solution is beefy enough, but we expect it to fall within the same 100-200MHz window we see with the existing models. Given the dual-die architecture, the 1950X almost certainly has 32MB of L3 cache, but AMD hasn't confirmed details.

The Threadripper 1920X features a 3.5GHz base and 4.0GHz boost. Intel hasn't released frequencies for its two competing chips, but it's pretty predictable. Intel's frequencies decline as the company adds more cores, and the 10C/20T Core i9-7900X features a 3.3GHz base. That means all of Intel's more well-endowed chips will have lower base frequencies. Both Threadripper processors will have higher base frequencies, partially offsetting some of Intel's IPC advantage. However, Intel's boost frequencies were one of the most surprising aspects of the Core i9 lineup. Intel's high core count models all boost to 4.5Ghz, and though we can't be sure of the unannounced turbos of the -7960X and -7920X, Intel might hold the advantage.

AMD's high-end models come without a cooler, and we expect the same trend for the Threadripper models. That's in-line with Intel's policy, too, so those extra costs are similar for both lineups. However, AMD employs solder while Intel stubbornly sticks to thermal paste, much to the ire of enthusiasts. We've tested the Core i9-7900X and found significant problems with the thermal paste, and we expect that to continue with Intel's higher core count models. AMD will likely have a higher TDP, though, so it will be interesting to see how it pans out when the chips hit our labs.

We don't know much else about the Threadripper lineup other than existing information. We do know the chips drop into the massive 4,094-pin TR4 socket with the X399 chipset. That's the same socket AMD uses for its EPYC data center lineup, and it dwarfs Intel's 2,066-pin LGA2066 socket. AMD has long held the advantage when it comes to motherboard pricing, and though the X299 platform can be characterized as anything but cheap, AMD's TR4 socket might lead to a similar price range. Simply put, larger sockets are more expensive.

The Threadripper lineup will come armed with quad-channel memory, which is a notable advantage Intel's X99 (and now X299) platform holds over the cheaper Ryzen 7 models. That should even the playing field. Intel also has its 18/36 Core i9-7980XE in reserve, which lends the company the core count advantage, and a much broader product stack with five models with varying core/thread counts. Neither of those advantages mean much if there is a comparable platform with much cheaper alternatives. AMD also has an overwhelming PCIe advantage with a total of 60 lanes, while Intel has varying PCIe allotments that are markedly lower across the board.

Ryzen 3

AMD also announced that the Ryzen 3 lineup is shipping on July 27. The Ryzen 3 1300X weighs in with four cores and four threads with a 3.5/3.7GHz base/boost, while the Ryzen 3 1200 features the same core/thread count at 3.1/3.4GHz. These chips will do battle with Intel in the low end of the market, which is perhaps the largest volume segment. Intel holds the advantage of its integrated graphics, which becomes more important as we travel down the stack to models that aren’t usually accompanied by discrete graphics cards.

However, we expect the same brutal pricing model from AMD, so competition will be pitched for the low end of the enthusiast segment. Intel hasn't readjusted its pricing on the low-end models, but now that AMD is squeezing the entire desktop spectrum, we could see a reaction.

Does Chipset-to-GPU Matching Matter?

Thomas Soderstrom — Wed, 04 Apr 2007 15:24:47 +0000

Should A Graphics Card Match The Chipset?

Ever since Nvidia released its first chipset, PC building newcomers have constantly questioned whether these must be matched to a like-branded graphics card. Traditionally, however, Nvidia chipsets have had no problems supporting ATI graphics cards, and the inverse became true when ATI later introduced its own chipsets.

But why would anyone do this? The answer was easy two years ago, when Nvidia's clear lead in AMD-compatible chipsets was at odds with ATI's clear lead in graphics performance; such were the days of 9700 Pro and FX 5800 graphics. But while today's chipsets seek to "lock in" graphics buyers by offering unique brand-only features such as SLI or Crossfire mode, the majority of buyers will never use these features. And though a larger number of users may wish to add multiple cards to support an increased number of monitors, chipset brand doesn't affect the use of dual-independent cards.

Putting aside normal expectations, such as the consistent compatibility between chipset and graphics products of battling brands or the loss of special chipset features such as SLI or Crossfire mode, little has been written to show the actual performance impact of such mix-and-match combinations. Certainly these companies have chipset optimizations that show preference towards own-brand graphics, don't they? If so, such optimizations could potentially invalidate chipset comparisons by making "identical configurations" favor the chipset brand from which the graphics card was selected!

But the benchmarking issue only links theory to reality: Many builders upgrade continuously rather than building from scratch. Should brand matching guide these upgrades?

The Test

This isn't about which manufacturer is leading in discrete graphics card or chipset performance. Instead, this review is about whether or not these companies employ "engineering trickery" to make their other products shine. The ones we know about, such as automatic overclocking of either the graphics card or PCI-Express interface, are BIOS options that should be disabled for any technology comparisons. Manual overclocking is available for the top products of both companies, which represents a better-performance option for anyone who really wants to push his or her system.

Thus, the rundown must include the top performing ATI and Nvidia graphics cards, as well as these companies' latest chipsets with automatic overclocking disabled.

ATI Graphics: Sapphire X1950XTX

While a few companies have produced overclocked versions, most X1950XTX graphics cards are identical right down to the sink. Today's X1950XTX "standard edition" card is brought to you courtesy of Sapphire.

Nvidia Graphics: Foxconn 8800GTX

As was the case with the competing graphics processor, most 8800GTX cards are identical save for a few special overclocked or water-cooled versions. This 8800GTX "standard version" was supplied by Foxconn.

Intel Platform: MSI P965 Platinum

The P965 chipset sets the baseline because it shouldn't have any graphics-brand specific optimizations. The award-winning MSI P965 Platinum was chosen for its performance leadership, keeping this older Intel mainstream chipset in the best possible light when comparing newer enthusiast parts from ATI and Nvidia.

ATI Platform: DFI ICFX3200-T2R/G

AMD’s purchase of ATI made platform selection easy, since the merger scared most of ATI’s motherboard partners into canceling plans for its new Intel chipset. Standing firmly in its commitment to ATI enthusiasts is DFI, a company frequently praised for its innovation.

The CrossFire Xpress 3200 has become the black sheep of the performance industry, first praised then ignored by Intel after rumors of the potential rivaling merger began to look realistic. This chipset’s market may finally blossom when AMD releases its long-awaited R600-based graphics cards, but until then the formerly-named "RD600" chipset is usually seen as nothing more than a slightly more modern alternative to Intel’s aging 975X.

One of the few things missing from TG Publishing’s earlier ICFX3200-T2R/G examination is a picture of the installation kit. Included are a quick installation guide, full manual, driver CD and windows RAID installation driver floppy, warranty card, round Ultra ATA and Floppy Disk Drive cables, six Serial ATA cables, two single-Molex to dual-SATA power adapters, the Karajan audio module with retaining clip, a port panel shield, a square case badge and a clear LanParty window decal.

Nvidia Platform : ECS PN2 SLI2+

Several Nvidia marketing partners have chosen the company’s reference motherboard, from high-end graphics card brands like BFG and ECS to budget-conscious motherboard brands like Biostar and yes, ECS. With so many brands selling the same product, picking one was as easy as grabbing the part located closest to the bench.

Test Setup

System Hardware
Socket 775 Processor	Intel Core 2 Duo E6700(Conroe 65 nm, 2.67 GHz, 4 MB L2 Cache)
Intel Chipset Motherboard	MSI P965 Platinum (Intel P965 Express Chipset)BIOS 1.3 (1-18-2007)
ATI Chipset Motherboard	DFI ICFX3200-T2R/G (ATI CrossFire Xpress 3200)BIOS 212 (2-12-2007)
Nvidia Chipset Motherboard	ECS PN2 SLI2+ (nForce 680i SLI)BIOS P25 (02/08/2007)
ATI Graphics Card	Sapphire X1950XTX 512M PCI-E D-DVI/VIVOATI Radeon X1950XTX (650 MHz), 512 MB GDDR4 (1550 MHz)
Nvidia Graphics Card	Foxconn GeForce 8800GTX, P/N : FV-N88XMAD2-ODNVIDIA GeForce 8800GTX (575 MHz), 768 MB GDDR3 (1800 MHz)
RAM	Corsair Dominator TWIN2X2048-8888C4DF2x 1024 MB DDR2-1111 (CL 4.0-4-4-12)
Hard Drive	Western Digital WD1500ADFD-00NLR1, Firmware : 20.07P20150 GB, 10,000 RPM, 16 MB cache, SATA/150
Power Supply	OCZ GameXStream OCZ700GXSSLI - 700W
System Software & Drivers
OS	Windows XP Professional 5.10.2600, Service Pack 2
DirectX Version	9.0c (4.09.0000.0904)
Platform Drivers	Intel Platform : INF 8.0.1.1010ATI SB600 6.11.100.1300NVIDIA Platform : nForce 9.53
Graphics Driver	ATI Catalyst 7.2Nvidia Forceware 97.92

Corsair Dominator 8888 was chosen for its ability to run 1066 MHz at CAS4, in hopes that all three motherboards would support this speed. EPP allowed automatic timing and voltage optimization using the "Expert" setting at a 1066 MHz data rate on the nForce 680i SLI platform. Because the Intel and ATI chipset motherboards do not support EPP, these settings were manually copied to the P965 Express and Radeon Xpress 3200 motherboards for the closest performance match.

In a perfect world this all would have been well, but part of the MSI P965 Platinum’s "class leading" performance appears to be that the company secretly changes any CAS 4 setting to CAS 3, at least according to CPU-Z. Before anyone becomes enraged about this form of "cheating," I’d like to commend MSI on producing the only board I’ve used that’s actually stable using this lower latency with the majority of CAS 4.0 DDR2 800 modules. MSI did not likely predict the quick development of DDR2-1066 CAS 4, a speed at which the CAS 3 "cheat" couldn’t be 100% stable. The board was "penalized" by using the easiest fix, setting latency to 5-4-4-12 rather than 4-4-4-12, and CPU-Z reported it running as set.

The DFI ICFX3200-T2R/G also had trouble running DDR2-1066, but trial and error found an eventual solution by raising the Northbridge PLL and Core voltages. Many other users have either succeeded or failed in their attempts to make this board’s performance stable at this memory speed, so your results may vary.

Benchmarks and Settings
3D-Games
Call of Duty 2	Version : 1.03Video Mode : 1024x768Anti-aliasing : offShadows : notimedemo demo2
F.E.A.R	Version : 1.0 RetailVideo Mode : 1024x768Computer : MediumGraphics : MediumTest Path : Options/Performance/Test Settings
Quake 4	Version : 1.2 (Dual-Core Patch)Video Mode : 1024x768Video Quality : defaultTHG Timedemo waste.maptimedemo demo8.demo 1 (1 = load textures)
Audio
Lame MP3	Version 3.97 Beta 2 (12-22-2005)Audio CD "Terminator II SE", 53 minwave to mp3160 kbps
OGG	Version 1.1.2 (Intel P4 MOD)Version 1.1.2 (Intel AMD MOD)Audio CD "Terminator II SE", 53 minwave to oggQuality : 5
Video
TMPEG 3.0 Express	Version : 3.0.4.24 (no Audio)fist 5 Minutes DVD Terminator 2 SE (704x576) 16:9Multithreading by rendering
DivX 6.2	Version : 6.2.2 (4 Logical CPUs)Profile : High Definition Profile1-pass, 3000 kb/sEncoding mode : Insane QualityEnhanced multithreadingno Audio
XviD 1.1.0	Version : 1.1.0 Beta 2Target qantizer : 1.00
Applications
Autodesk 3D Studio Max	Version : 8.0Characters "Dragon_Charater_rig"rendering HTDV 1920x1080
Synthetics
PCMark05 Pro	Version : 1.1.0System, CPU and Memory TestsWindows Media Player 10.00.00.3646Windows Media Encoder 9.00.00.2980
SiSoftware Sandra 2005	Version 2005.7.10.60CPU Test = Arithmetic, Multimedia BenchmarksMemory Test = Bandwidth Benchmark

Benchmark Results

Many reviews have shown 8800GTX beating up the X1950XTX in the majority of games, but there are a few that still favor the older ATI card. Let's see what effect chipset selection.

3D Games

Starting off with one game test, the X1950XTX can win, while the 680i SLI leads both Intel and ATI chipsets using the competitor's card. So much for ATI chipset to VPU optimizations.

F.E.A.R. put Intel's P965 in the lead using the 8800GTX, flying in the face of any perceived "chipset-to-GPU" optimizations on Nvidia's behalf.

Once again Intel's chipset gives Nvidia's graphics card more performance than Nvidia's chipset. Perhaps the credit should instead go to MSI, the maker of this P965 Express motherboard.

Applications

Chipset-to-graphics performance optimizations don't have to be restricted to games alone, as hidden benefits could include such sneaky things as memory-path optimization. Such tricks would surely show up in applications.

3D Studio Max favors the Intel chipset slightly, or is it the X1950XTX? It seems odd that both the ATI/ATI and Nvidia/Nvidia combinations fall short.

Audio

The 680i SLI wins both audio benchmarks, but doesn't seem to care whether it's using the ATI or Nvidia graphics card.

Video

The P965 Express chipset via MSI's P965 Platinum takes a fairly significant leap in DivX, and retesting proved this lead consistent. Chipset to specific-brand GPU optimizations certainly aren't seen here, as the 8800GTX leads by one second using both ATI and Nvidia chipset motherboards.

P965 Express video encoding domination continues in XviD, with the CrossFire Xpress 3200 lagging, but what about graphics-specific chipset optimizations? Intel favors Nvidia while Nvidia favors ATI, which offers more evidence against graphics-specific chipset optimization.

Synthetic

In PC Mark's CPU test, Intel favors Nvidia graphics, Nvidia favors ATI graphics and ATI favors Nvidia graphics while the latter two results show a possible bias in favor of competing technology! This peculiar coincidence is almost ironic.

The 680i SLI takes PC Mark's RAM test, but once again wins using a competing graphics card.

Nvidia's chipset again leads in PC Mark's System test, while this time it has a Nvidia card.

Synthetic, Continued

Sandra's CPU Arithmetic and Multimedia benchmarks show the MSI Intel chipset motherboard leading with either the ATI or Nvidia card, and Nvidia's chipset doesn't seem to mind at all which graphics card it uses. The ATI chipset usually favored ATI graphics, but never really stood out in a way that might point to a cheat.

Big numbers for Intel's memory controller put both Nvidia and ATI chipsets to shame, but speaking of the latter two, neither showed a consistent preference for configurations containing own-brand graphics cards.

Conclusion

The benchmarks consistently show inconsistency between graphics and chipset performance, so brand matching of graphics cards to chipsets doesn't seem to matter at all. If you're not concerned about SLI or Crossfire, your card selection is as simple as picking the best card available for the money, while you can rest easy in the knowledge that mixing these imparts no performance penalty. This is great news for anyone who upgrades one part at a time.

Another thing that stood out was Intel's greater number of wins using its "mainstream" chipset, which was especially noticeable in Sandra's Memory bench, even though the particular motherboard chosen was forced to use CAS 5 while the others used CAS 4 latency setting. As Intel gets friendlier with Nvidia following the ATI/AMD merger, Nvidia might be persuaded to allow SLI mode for Intel chipsets, but only time will tell.

Core 2 processor buyers looking forward to an R600-based Crossfire solution are currently limited to Intel's ancient 975X and ATI's yet-to-mature CrossFire Xpress 3200 chipset. However, they too can also rest assured that their "old" PCI-Express graphics card will carry them through to release day, regardless of whether the carried-over card is based on ATI or Nvidia technology.

GeForce 8800: Here Comes the DX10 Boom

Darren E. Polkowski — Wed, 08 Nov 2006 19:59:37 +0000

Holy Hardware Batman!

GeForce 8800GTX is head and shoulders above the competition.

We have been hearing about DirectX 10 hardware and the miraculous advantages it has over DX9. We have even seen the screenshots of games in development. However, until today, the hardware has been lacking. Earlier today we wanted to whet your appetites for Direct3D (D3D) hardware and Nvidia delivers the goods first. We are pleased to announce the arrival of DirectX 10 compliant hardware in the form of Nvidia's GeForce 8800GTX and 8800GTS.

As our DX10 preview article concluded, a unified architecture will get more gains out of the shader units as they can be utilized more efficiently than a fixed function layout. Ushering in the new era of computer graphics is the GeForce 8800GTX with 128 unified shader units and the GeForce 8800GTS with 96 shader units. Long gone are the days of pipelines, finally. Without further ado, let's look inside the mouth of the beast and see how this thing works.

slide show: GeForce 8800GTX/GTS

More Information on DirectX 10

What Direct3D 10 Is All About
Today's launch of the Nvidia GeForce 8 series marks the advent of next-generation graphics. What can we expect from graphics makers with respect to DirectX 10 hardware?
The New Graphics
A tale of Direct X 10, and rumors of the hardware to drive it. While the demand for Direct X 9 hardware is not slipping, and more graphics cards are constantly being launched, there is much interest in this new standard and the hardware that will support it.
The Graphics State Of The Union
Tom's Hardware graphics presidente Polkowski is concerned about the 3D arms race. Power and heat dissipation are skyrocketing, but external graphics boxes could eliminate the imminent need for 1,000 W power supplies.

Technology:
Nvidia Tightens Up Midrange Performance With The $300 GeForce 7950GT
The GeForce 7900GS is Nvidia's New Mid-Range
Can SLI in a Notebook Beat Desktop Graphics?
ATI's Radeon X1950XTX: You Say You Want A RAM Revolution

United We Stand

Worth twice its weight in gold, a wafer with 80 graphics processing cores can deliver about twice the performance than GeForce 7900GTX (G71). A 681-million transistor count makes for a large silicon footprint, but when asked about its size, CEO Jen-Hsun Huang replied: "If my engineers said that they could double performance by doubling the amount of silicon used, I would have told them 'go for it!'"

As time has shown, doubling does not mean double performance but Nvidia seems to have struck the right balance of technology advances with silicon engineering and implementation.

Staying close to DX10 specifications, GeForce 8800GTX and 8800GTS fully comply with the DX10 standard with Shader Model 4.0, various data storage and transmitting specifications, Geometry Shaders and Stream Out. While you have seen what DX10/D3D10 compliant hardware should operate, let's look at how Nvidia gets the job done.

To start, Nvidia deviated from the fixed function design that the industry had been using for the past 20 years in favor of a unified shader core.

We have shown similar images in the past demonstrating the trend towards more pixel shading. Nvidia as well understands this trend and moved to balance the utilization needs by implementing unified shaders running threaded data streams to maximize efficiency and performance.

Nvidia said: "The GeForce 8800 design team realized that extreme amounts of hardware-based shading horsepower would be necessary for high-end DirectX 10 3D games. While DirectX 10 specifies a unified instruction set, it does not demand a unified GPU shader design, but Nvidia GeForce 8800 engineers believed a unified GPU shader architecture made most sense to allow effective DirectX 10 shader program load-balancing, efficient GPU power utilization and significantly improved GPU architectural efficiency." This logically makes the most sense as pointed out in our Direct3D 10 Preview.

SIMD Heaven: 128 Streaming Processors

Click to see a larger version of this image.

The processor core itself operates at a frequency of 575 MHz for the GeForce 8800GTX and 500 MHz on GeForce 8800GTS. While the rest of the core runs at 575 MHz (or 500 MHz), the shader core has its own independent clock generator. GeForce 8800GTX runs at 1,350 MHz and the 8800GTS' clock speed is 1,200 MHz.

"Streaming Processor" is the term given to each shader core unit. The GeForce 8800GTX has 16 sets of eight streaming processors in a block. A total of 16 blocks makes up the entire number of 128 SPs. Like ATI's design in R580 and R580+ with its Pixel Shader units, Nvidia stated that more units can be added to future designs and some can be taken away. This can be seen in the implementation of 96 streaming processors in GeForce 8800GTS.

Click to see a larger version of this image.

More Details

The previous issues with Nvidia not being able to do AA with HDR are history. Each ROP supports frame buffer blending. This means that both FP16 and FP32 render targets can be used with multi-sample antialiasing. Under D3D10, eight Multiple Render Targets can be utilized in conjunction with new compression technologies to accelerate color and Z processing in the ROPs.

The GeForce 8800GTX can fill 64 textures per clock cycle, and at 575 MHz, it can serve up a maximum of 36.8 billion textures per second (GeForce 8800GTS = 32 billion/sec). The GeForce 8800GTX has 24 ROPs, and when running at a core frequency of 575 MHz, the card has a peak pixel throughput of 13.8 Gpixels/sec. Similarly, the GeForce880GTS version has 20 ROPs and therefore has a peak fill rate of 10 Gpixels/sec at 500 MHz.

Nvidia GeForce Reference Specifications
	8800GTX	8800GTS	7950GX2	7900GTX	7800GTX 512	7800GTX
Process Technology (nm)	90	90	90	90	110	110
Processor Core	G80	G80	G71	G71	G70	G70
Number of Processors	1	1	2	1	1	1
Number of Transistors per Processor (Millions)	681	681	278	278	302	302
Vertex Frequency (MHz)	1350	1200	500	700	550	470
Core Frequency (MHz)	575	500	500	650	550	430
Memory Clock (MHz)	900	600	600	800	850	600
DDR Rate (MHz)	1800	1200	1200	1600	1700	1200
Vertex Shaders (#)	128	96	16	8	8	8
Pixel Shaders (#)	128	96	48	24	24	24
ROPs (#)	24	20	32	16	16	16
Memory Interface (bit)	384	320	256	256	256	256
Frame Buffer Size per Processor (MB)	768	640	512	512	512	256
Memory Bandwidth (GB/sec) per processor	86.4	48	38.4	51.2	54.4	38.4
Vertices/Second (Millions)	10800	7200	2000	1400	1100	940
Pixel Fill Rate (# ROPs x clk) in Billions/sec	13.8	10	16	10.4	8.8	6.88
Texture Fill Rate (# pixel pipes x clk) in Billions/sec	36.8	32	24	15.6	13.2	10.32
RAMDACs (MHz)	400	400	400	400	400	400
Bus Technology	PCI Express	PCI Express	PCI Express	PCI Express	PCI Express	PCI Express

Check out the GeForce 8800GTX/GTS slide show.

If you were scratching your head at the memory bus width, here is your answer. From the core logic image on the previous page, you can see that six memory partitions exist on a GeForce 8800 GTX GPU. Each of these provides a 64-bit interface to memory, yielding a 384-bit combined interface width. The 768 MB of GDDR3 frame buffer memory is attached to a memory subsystem that utilizes a high-speed crossbar design, similar to GeForce 7x GPUs. This crossbar supports DDR1, DDR2, DDR3, GDDR3 and GDDR4 memory.

The GeForce 8800GTX uses GDDR3 memory, with a default clock speed of 900 MHz (GTS version is clocked at 800 MHz). With a 384-bit (48 byte-wide) memory interface running at 900 MHz (1800 MHz DDR data rate), frame buffer memory bandwidth is very high at 86.4 GB/sec. With 768 MB of frame buffer memory, far more complex models and textures can be supported at high resolutions and image quality settings.

Test Setup

The GeForce 8800GTX is longer than a Radeon X1950XTX and most other cards on the market. In fact it is 11" long tip to tail. The 8800GTS and other cards on the market are only 9.5" long.

As a comparison with the benchmarks we have produced throughout the entire 2006 calendar year, we ran our initial tests on an AMD Athlon FX-60 based system. We will have subsequent results showing multi-GPU performance. Additionally, we want to show what these new cards can do when the CPU limitations are lifted when testing under low resolution and image quality. (Stay tuned for that...)

System Hardware

Processor(s)

AMD Athlon 64 FX-60

2.6 GHz, 1.0 GHz HT-Link, 1 MB L2 cache

Platform

Nvidia : Asus AN832-SLI Premium

Nvidia nForce4 SLI, BIOS version 1205

RAM

Corsair CMX1024-4400Pro

2x 1024 MB @ DDR400 (CL3.0-4-4-8)

Hard Drive

Western Digital Raptor, WD1500ADFD

150 GB, 10,000 rpm, 16 MB cache, SATA150

Networking

On-Board nForce4 Gigabit Ethernet

Graphics Cards

ATI Radeon X1950XTX 512 MB GDDR4

650 MHz Core

1,000 MHz Memory (2.00 GHz DDR)

Nvidia Based Cards :

Nvidia GeForce 8800GTX 768 MB GDDR3

575 MHz Core

1,350 MHz Streaming Processor

900 MHz Memory (1.80 GHz DDR)

XFX GeForce 8800GTS 640 MB GDDR3

500 MHz Core

1,200 MHz Streaming Processor

800 MHz Memory (1.60 GHz DDR)

Nvidia GeForce 7900GTX 512 MB GDDR3

675 MHz Core

820 MHz Memory (1.64 GHz DDR)

Power Supply

PC Power & Cooling Turbo-Cool 1,000W

CPU Cooler

Zalman CNPS9700 LEDSystem Software & Drivers

Microsoft Windows XP Professional 5.10.2600, Service Pack 2

DirectX Version

9.0c (4.09.0000.0904)

Graphics Driver(s)

ATI - Catalyst 6.10 WHQL

Nvidia - Forceware 96.94 Beta

slide show : GeForce 8800GTX/GTS

Benchmarks Results

3DMark05

In 3DMark we tested at all resolutions with antialiasing and anisotropic filtering disabled in the first run. The second time around we enabled the image quality with 4xAA and 8xAF. Nvidia has the clear upper hand in 3DMark05. The GeForce 8800GTX can produce the same score at 2048x1536 as the ATI Radeon X1950XTX can in the default 1024x768 run. All I can say is: "Impressive, impressive, impressive."

Doom 3

Doom 3 is typically dominated by Nvidia as its core design was matched for this game. However, ATI has taken the title with single core designs and lower overhead drivers.

This is where we encounter the first of some CPU limitations as the low resolution and image quality settings bottleneck around 126 frames per second. The ATI card can produce higher frames per second under the current system configuration. The reason is the driver. ATI has been producing drivers with less CPU overhead associated with their design. This means that the CPU is utilized less and can output more performance for the graphics subsystem.

The winner overall is clearly the new 8800 series cards. Looking at the results at all resolutions, the new DX10 cards push past the Radeon X1950XTX from 1280x1024 and up.

F.E.A.R

Cards with Nvidia processors are particularly well geared for F.E.A.R. across all disciplines. Once again, we see the low driver overhead from ATI's drivers come into play. We are certain we will see different results under a faster system setup but for those of you with slower systems, this gives you an insight into how G80 cards will do on your system. You will notice that once the first 1024x768 test is over, G80 kills the Radeon X1950XTX. The GTX is a monster. No matter what we throw at the GeForce 8800GTX, it takes it and gives us more than 40 frames per second.

Oblivion

Elder Scrolls IV: Oblivion has been our ultimate torture test. This is the only title with which the GeForce 8800GTX cannot produce more than 30 frames per second in a test. However, that is when the resolution is maxed out at 4 MPixels per frame (2560x1600) and every feature is enabled in the game. Scaling back to 2048x1536 with all of the bells and whistles turned on plays above 30 fps in the outdoor scenes, which is certainly impressive.

Here are two images from our testing. The first is the image we get on the ATI X1950XTX at 2560x1600.

Click to see a larger image

The second (below) is with the 8800GTX with the same settings.

Click to see a larger image

This image is far superior in quality compared to that of the ATI. It looks like the tables have turned in that department. That brings us to other advances Nvidia has made with G80.

New Image Quality Settings

Here is a listing of the new image quality enhancements in G80 based hardware.

Nvidia showcased a lot of cool things at Editor's Day this year. Beyond the new DX10 graphics cards, they showed us some tools and features that will be enabled on G80 based cards. The first of these is an image quality functionality they have patented called Coverage Sampled Antialiasing (CSAA).

This new version of antialiasing uses a Boolean coverage at 16 sub-samples. According to Nvidia it can compress "redundant color and depth/stencil information into the memory footprint and bandwidth of four or eight multisamples." This new quality level works more efficiently by lessening the storage cost per sample. If a CSAA selection cannot work within a particular application (game), then the driver will revert back to a traditional AA method.

Here is a listing of the coverage sample cost reductions (click for a larger image).

Here is an example of the image quality improvement of 16xQ CSAA over traditional 4xAA. The footprint is a little more than 4xMSAA but produces a better result.

Click for a larger image.

Beyond Graphics

Before closing out this review, we would like to leave you with two more aspects of graphics cards that have been in development over time and are becoming more prominent. The first is video playback. Under the old regime of GeForce 7, ATI Radeon X1900 cards dominated the video playback quality arena. This reign comes to an end with the advent of unified shaders with a dedicated Pure Video core.

With smarter algorithms and 128 shaders to pick from, the GeForce 8800GTX was able to score a 128 out of 130 points in HQV. We will be doing some updated image quality test of our own as we just visited this topic, so stay tuned for more on this subject.

Last but not least on our tour of G80 enabled treats is what Nvidia calls CUDA. This was one of the things that blew me out of the water. For years there have been enthusiasts and academicians looking for ways to do more on powerful parallel processors. When a Beowulf cluster was out of reach for the budget conscience, people started looking to graphics cards for more horsepower.

The problem with using graphics is that it is great at parallel processing but can be lousy at high-branching problems. This is where a CPU is better. To top things off, one has to program shaders like a game developer just to get the problems solved on a graphics card. Nvidia once again pulled out a trump card with Compute Unified Device Architecture or CUDA.

Here is a breakdown of what CUDA can do with a fluid simulation.

Nvidia created a C+ compiler that scales with GPU horsepower (like the 96 SPs in 8800GTS vs. 128 in 8800GTX). Now programmers have the ability to create programs that scale on both the CPU and the graphics processor. Protein folding at home and other applications requiring massive amounts of computational power and precision will fall in love CUDA. Not only can CUDA help invent programs to solve the world's ills but it can be used to blow things up better or simulate other things like volumetric fluids, cloth and hair. Effects physics and even other game play features could be potentially ported back to the graphics processor via CUDA.

An entire SDK will be made available to those whom it should go to. If you are looking for a cure for cancer or ways to help mankind, Nvidia said they are willing to help you out.

Check out the GeForce 8800GTX/GTS slide show.

Conclusion

Amazing! That is the long and short of what we have covered today. To keep with tradition, Nvidia has these new cards available right now. There are e-tailers willing to sell them to you if you wish to spend the suggested price of $599 for GeForce 880GTX and $449 for GeForce 8800GTS. After seeing the results and what will be in store for the future, let us leave you with one more thought: the games are just around the bend, and not only will DX10 hardware play them when they arrive, but more importantly, they can play your games better NOW.

Here is an example of what a supermodel can look like when rendered in DX10. Meet Adrianne Curry in all her Direct3D 10 splendor. (Click for a larger image)

Author's Opinion

I am impressed by Nvidia's implementation DX10/D3D10. Seeing Crysis running in real-time and seeing some of the demos that Curtis Beason and his team put together is amazing. The implementation of CUDA into the picture makes the graphics card something much more than a frame rendering device. C+ compiled programs running on the CPU in conjunction with the parallel computing power of a versatile shader core with Stream Out is beyond comprehension as the possibilities are endless. I can't wait to see the community's implementation of such a wonderful tool.

G80 leaves me wanting more. Why, might you ask? The only thing lacking is the CONTENT! Come on software developers... we want to see the DX10 games NOW!

Read on: More on DirectX 10

What Direct3D 10 Is All About
Today's launch of the Nvidia GeForce 8 series marks the advent of next-generation graphics. What can we expect from graphics makers with respect to DirectX 10 hardware?

The New Graphics
A tale of Direct X 10, and rumors of the hardware to drive it. While the demand for Direct X 9 hardware is not slipping, and more graphics cards are constantly being launched, there is much interest in this new standard and the hardware that will support it.

The Graphics State Of The Union
Tom's Hardware graphics presidente Polkowski is concerned about the 3D arms race. Power and heat dissipation are skyrocketing, but external graphics boxes could eliminate the imminent need for 1,000 W power supplies.

Related Links:

slide show: GeForce 8800GTX/GTS

Buyer's Guides:
The Best Video Cards For Your Money: November 2006

Performance Charts (Overview on all available graphics solutions):
Interactive VGA Charts Update October 2006

News on TG Daily:
Nvidia fires away with new Geforce 8800 GPU

Barrier Eliminators: X1900GT and X1950XTX from Sapphire

Thomas Soderstrom — Tue, 07 Nov 2006 13:36:36 +0000

Quantity Takes On Quality

The idiom "you get what you pay for" definitely applies to the video card market, as buyers expect noticeable increases in performance with every significant increase in price. Game fanatics speak of frame rates, but ignore those "high frame rates" available from low-priced cards when used at low resolutions and reduced visual effects. Meanwhile, HD movie fans look to good graphics processors as a way to display high-resolution video smoothly and in fine detail. Like Marie Silvia has said, "it's all about the quality!"

Sapphire's X1950XTX and X1900GT hit the pricing sweet spots for upper and lower enthusiast markets, while EVGA's Superclocked 7600GT stands in for comparison. Let's take a look at how these technologies cope as we crank up the quality levels and resolutions.

Brand Identity?

The drive to bring new models to the market quickly has forced most of ATI's production partners to use reference PCB designs, but that doesn't necessarily mandate uniformity; Besides overclocking of some "premium" models, companies often differentiate through package design, features, accessories and price.

A colorful box might be the primary concern for local-level retailers, but enthusiasts just want to get at what's inside. Most noticeable on Sapphire's X1900GT is ATI's "old-school" small fan reference cooler, which has been tweaked to operate at lower speeds and noise levels on this reduced 36-pixel shader 575 MHz processor. Its super-cool X1950XTX counterpart is also based on a reference design, where huge copper sinks and heatpipes chill the 48-pixel shader 650 MHz processor.

Sapphire's X1900GT and X1950XTX support both AVIVO and VIVO, which are similar-sounding technologies that should never be confused. AVIVO encompasses ATI's latest video quality enhancements and namesake hardware-assisted video converter software, while VIVO simply means Video In/Video Out. Though most modern cards support TV-Out (the VO part of VIVO), only selected models such as these support TV-In. ATI's Rage Theater chip encodes Video Inputs, which can be captured using Windows XP's included software (Windows Movie Maker), third-party applications or possibly ATI Multimedia center, if you already have a licensed MMC product installed.

Sapphire bundles both cards with a manual, driver CD, a reduced-feature version of Cyberlink PowerDVD, a trial pack of Cyberlink PowerDirector 4, a four-pin Molex to six-pin PCI-Express supplemental power connector, Video In/Video Out splitter cable (supporting S-Video and Composite Video), component video break-out splitter, S-Video cable, Composite Video cable and two DVI-I to VGA adapter blocks.

The X1950XTX comes with The DaVinci Code PC-DVD game, while the X1900GT comes with a four-game DVD trial bundle and coupon code to keep one game of your choosing. Titles for the X1900GT game bundle are Brother In Arms: Road to Hill 30, Prince of Persia: Warrior Within, Richard Burns Rally and Tony Hawk's Underground 2.

Other than the pretty box and free game, Sapphire competes with other ATI partners on price. Starting at $200 for its X1900GT and $400 for its X1950XTX, Sapphire is among the most reasonably priced brands on the Web.

Test Setup

System Hardware
Processor	Intel Core 2 Duo E6700, 2.66 GHz, 1066 FSB, 4 MB Cache
Motherboard	Abit AB9-Pro V1.00, BIOS 1.4
RAM	Corsair CMX512-8000UL2x 512 MB, CAS 4-4-4-8
Hard Drive	Western Digital Raptor WD1500ADFD150 GB, 10,000 RPM, 16 MB Cache, SATA 150
Networking	Marvell Yukon Gigabit PHY, enabled, offline
Graphics Card 1	Sapphire X1950XTX 512M PCI-E D-DVI/VIVOGPU : ATI Radeon X1950XTX (650 MHz)RAM : 256 MB GDDR3 (1550 MHz)
Graphics Card 2	Sapphire X1900GT 256M PCI-E D-DVI/VIVOGPU : ATI Radeon X1900GT (575 MHz)RAM : 256 MB GDDR3 (1200 MHz)
Graphics Card 3	EVGA 7600GT CO SuperclockedGPU : Nvidia GeForce 7600GT (600 MHz)RAM : 256 MB GDDR3 (1560 MHz)
Power Supply	OCZ GameXStream OCZ700GXSSLI - 700W
System Software & Drivers
OS	Windows XP Professional 5.1.2600, Service Pack 2
DirectX Version	9.0c (4.09.0000.0904)
Platform Driver	Intel INF 8.1.0.1006
Graphics Drivers	ATI Catalyst 6.9Nvidia Forceware 91.47
Benchmarks and Settings
3D-Games
Black & White 2	Version : Patch 1.1Video Modes : 1024x768, 1280x1024, 1600x1200Video Quality : Medium, High with 4xAATest : FRAPS 60-second - Intro Island Flyby
FEAR	Version : 1.0 RetailVideo Modes : 1024x768, 1280x960, 1600x1200Computer : MaximumGraphics Card : Medium, Maximum (Softshadows, 4xAA, 16xAF)Test Path : Options/Performance/Test settings
Oblivion	Version : 1.0.228 RetailVideo Modes : 1024x768, 1280x1024, 1600x1200Visual Settings : Bloom, HDR, 4x AATest : THG custom Indoor/Outdoor game saves
Quake 4	Version : 1.2 (Dual-Core Patch)Video Modes : 1024x768, 1280x1024, 1600x1200Video Quality : Medium, Ultra with 4xAACommand Line : timedemo demo8.demo 1 (1 = load textures)
Video Quality
HQV DVD	Catalyst : 3:2 Pulldown Detection + Auto QualityForceware 1 : Standard SettingsForceware 2 : Edge Enhance. + Noise Red. + Inverse telecine

Note that both high quality and standard quality game settings are listed in the chart above, with Oblivion tested at three quality settings.

Video Quality : Movies

Enabling high-quality DVD viewing under ATI Catalyst Control Center is as easy as turning on automatic 3:2 pull-down detection, which makes us wonder why this option isn’t enabled by default. Nvidia Forceware requires manual selection to best suit the card’s purpose. Analog sliders allow settings between maximum and minimum enhancement, where the driver defaulted to the lowest setting.

Designed for testing high-end home theater displays, the HQV DVD from Silicon Optix assigns numeric values to qualitative observations. Because manual mode selection is required for Forceware drivers, the 7600GT was tested at both its base setting and with full enhancements/inverse telecine enabled.

Silicon Optix HQV DVD Performance Results
Test	X1950XTX (Auto)	X1900GT (Auto)	7600GT (Base)	7600GT (Optimized)
Color Bar (10)	10	10	10	10
Jaggies 1 (5)	5	5	3	3
Jaggies 2 (5)	3	3	3	3
Flag (10)	5	5	5	5
Picture Detail (10)	0	0	0	10
Noise Reduction (10)	10	10	5	5
Motion Adaptive NR (10)	5	5	5	5
3:2 Detection (10)	5	0	0	10
Film Cadence (8 tests x5)	40	40	0	40
H. Text Crawl (10)	10	10	10	0
V. Text Crawl (10)	10	10	10	0
Totals (130 max)	103	98	51	91

A simple explanation of scoring differences is more practical than 100 medium resolution screen shots :

Color Bar Test : All three cards showed the smallest bars well, without flickering.

Jaggies Pattern 1 : Scan lines first appears on the X1900GT and X1950XTX after the bar reached 10 degrees to horizontal, while the 7600GT exhibited scan lines as the bar first approached 10 degrees to horizontal. The actual difference is approximately two degrees, but this is the barrier at which the test defines three or five points.

Flag Test : The X1900GT and X1950XTX showed a softened background without noticeable interlacing. The 7600GT also exhibited the same behavior using base settings, while its optimized settings sharpened the background and introduced very slight scan lines.

Picture Detail Test : All three cards initially exhibited a softened background, but selecting optimum settings allowed the 7600GT to pass.

Noise Reduction Test : The X1900GT and X1950XTX eliminated analog noise from the analog-to-digital converted video segments. The 7600GT left some noise in background colors (sky) using both driver settings.

Motion Adaptive Noise Reduction Test : The X1900GT and X1950XTX showed minor blur around the moving object (a roller coaster). The 7600GT exhibited background noise (both settings).

3:2 Detection : The X1950XTX eliminated moiré in slightly less than half a second. Though based on the same core technology, the X1900GT failed to perform 3:2 detection promptly. The 7600GT showed constant moiré with inverse telecine disabled (base setting) and no moiré after enabling it.

Film Cadence Tests : The X1900GT and X1950XTX passed all eight tests with no moiré pattern or interlacing problems. The 7600GT exhibited a moiré pattern on the newspaper using base settings, but eliminated it after enabling inverse telecine.

Horizontal Text Crawl : The X1900GT and X1950XTX showed smoothly scrolling text with well defined edges and no interlacing in the background image. The 7600GT showed smoothly scrolling text without background interlacing at its base setting, but added noticeable background interlacing after optimizations were enabled.

Vertical Text Crawl : The X1900GT and X1950XTX displayed smoothly scrolling text on a clear background, as did the 7600GT at its base setting. Enabling advanced 7600GT driver features introduced a significant amount of minor background noise, with a loss in detail.

The 7600GT exhibited well-defined edges surrounding the text (no blurring) in both scrolling text tests, but showed significant noise in the text color. The HQV DVD scorecard does not account for such behavior, so no additional points were forfeited.

System Performance: Movies

Both ATI and Nvidia would like you to believe that HD Content will suck your CPU dry, and you really need a card with the respective company's processor to enable smooth playback via hardware acceleration.

Using Power DVD 6, all three cards consumed approximately 4% of available CPU cycles while playing DVD video with hardware acceleration disabled; the percentage dropped to around 1% with hardware acceleration enabled. Of course, 480p MPEG2 content isn't HD.

We downloaded the "1080p" movie trailer "The Rules of Attraction" from Microsoft's Windows Media Showcase, but were dumbfounded when all three cards again reached the same CPU cycle consumption at approximately 14% in Windows Media Player, regardless of whether or not WMV Acceleration was enabled in the Catalyst driver. Perhaps we should try H.264?

Apple's Quick Time HD Gallery offers a wide variety of H.264 content. We chose the documentary clip "One Six Right" in true 1080p format, only to find all three cards once more using the same amount of CPU cycles to play content, this time approximately 22%. Given the identical results, it would seem as though none of the downloadable HD formats require or use hardware acceleration, and further testing will require the procurement of an HD-DVD or BRD drive. If the DVD playback results showing a 75% reduction in CPU use indicate what to expect, we might be in for a treat!

Video Quality: Games

Catalyst High-Quality 3D Settings
Smoothvision HD AA	Application Preference
Smoothvision HD AF	Application PreferenceHigh Quality AF Enabled
Catalyst AI	Standard
Mipmap Detail Level	Quality
Forceware High-Quality 3D Settings
Anisotropic Filtering	Application Preference
AF MIP Filter Optimization	Off
Anisotropic Sample Optimization	Off
Anti-Aliasing Settings	Application Preference
Conformant Texture Clamp	Use Hardware
Extension Limit	Off
Force Mipmaps	None
Gamma Correct Anti-Aliasing	Off
Hardware Acceleration	Single Display Performance Mode
Ignore OpenGL Error Reporting	Off
Negative LOD Bias	Allow
OpenGL Threading	Auto
Texture Filtering	High Quality
Transparency Anti-Aliasing	Off
Trilinear Buffering	Off
Triple Buffering	Off
Vertical Sync	Application Preference (all apps set off)

I'm not personally much of a "gamer" in the popular sense, preferring the fast action of racing games to RPG and FPS. As such, I was unable to reach "the perfect scenes" for screenshots, but grabbed what I could instead. Though most X1950XTX, X1900GT and 7600GT screenshots looked comparable, one scene stood out:

At maximum quality settings in FEAR, the 7600GT (above) displayed the building on the left normally, but with little detail in the right-side roof or rear parking lot. Compare the X1950XTX, which displays a strange moiré affect in the building, but with greater detail in the right-side roof and rear parking lot. This is the only place I found where this visual effect occurred, but I've never seen it mentioned before. It affects both the X1900GT and X1950XTX.

The sample above would appear to show ATI with stronger AF and Nvidia with more accurate AA results. It's difficult to make any definitive statements regarding ATI vs. Nvidia in the remaining shots, but feel free to browse the photo gallery for argument's sake.

3D Performance

Black & White 2

Focusing specifically on the impact high-quality settings have on frame rates allows one to determine each card's maximum "playable" quality level. We begin with Black & White 2, using the default medium settings, then the maximum settings plus 4x anti-aliasing (AA-High).

The X1950XTX appears to be the minimum card needed to play Black & White 2 at high resolutions and quality. Buyers on tight budgets will find adequate performance from the X1900GT at high quality and medium resolutions, but 7600GT might want to upgrade.

F.E.A.R

In Fear, we chose Maximum System/Medium Graphics as a base line, moving up to Maximum Graphics (4xAA/16xAF with Soft Shadows enabled) for the quality/performance comparison.

All three cards can play at maximum quality up to 1280x960 in FEAR, but the 7600GT drops out of the race at 1600x1200 pixels.

Quake 4

Quake 4 defaulted to medium settings. We used the "ultra" setting and added 4xAA for high-quality performance comparisons.

Quake 4 presents a unique scenario in "the way it's meant to be played," as the only game tested to put the X1900GT and 7600GT in the same league, at least when using default settings. Ultra-quality mode puts things back into perspective, with the X1900GT surviving to 1600x1200 resolution.

Oblivion

Oblivion supports HDR lighting or anti-aliasing using WHQL-Certified drivers, so we tested both. All distant rendering settings were turned on, making the outdoor environment particularly brutal for graphics processors.

The X1950XTX and X1900GT take tremendous performance hits when HDR rendering is enabled, but the 7600GT seems barely affected. Moreover, both 256 MB cards were completely unsuited for the outdoor environment, especially at resolutions beyond 1024x768, which thus justifies the X1950XTX's higher price.

The Cost Of Quality

It's unfortunate that none of our downloaded HD content was able to make good use of hardware-accelerated decoding, because it seems a forgone conclusion that buyers will need something better than onboard graphics to play HD-DVD and BRD movies smoothly. At least the video quality tests proved useful, with ATI winning handily.

In games, the X1900GT proved adequate for applying high quality settings at medium to high resolutions, whereas the older 7600GT often couldn't compete. The real shocker is that this particular 7600GT model, EVGA's Superclocked Edition, is priced similarly to the X1900GT. Surely we see the value in the X1900GT, as increased capacity for high-quality visual settings adds nothing to its price.

So the X1900GT is a good value, but some people will certainly want more. Those with deep pockets will appreciate that the X1950XTX is up to twice as fast at twice the price, giving it similar "bang for the buck."

Author's Opinion

As a fan of fast-paced racing games, I don't often have the time to notice edge aliasing, though anisotropic filtering may keep the distant road surface well defined. My largest complaint has always concerned deeply textured and raised surfaces such as stone walls and railroad tracks, which always appear to be painted onto a flat surface. Even as the X1950XTX provided exceptional performance improvements at the highest quality levels, we're several generations from a graphics card that can provide the added detail I desire.

ATI's Radeon X1950XTX: You Say You Want A RAM Revolution

Darren E. Polkowski — Wed, 23 Aug 2006 19:15:17 +0000

From EDO-DRAM To GDDR4

Boosting the performance of computer graphics has always been an art and there are various ways to do it. In recent years, we have seen larger graphics processors that host more pixel pipelines as well as fragmented designs, such as ATI's Radeon X1000 series; and there have been attempts to bolster 3D performance by deploying graphics cards in dual and quad configurations. The classic way is through innovations in design and the introduction of modern approaches and components and ATI wants to prove again that this approach still works. Welcome the Radeon X1950XTX.

Looking back in history we can reminisce to the beginning of computer graphics as we know it today. About 10 years ago, 3D graphics started to gain success. A drop in the price of EDO (Extended Data Out) memory chips made it affordable for add-in graphics cards makers to get their product out to the mainstream.

From there, the next logical step for progress comes from new applications that can utilize the upgraded hardware. Many different application programming interfaces (APIs) have been developed, but they continue to mature as newer ideas are hatched in the minds of developers and engineers to make everything quicker, better, faster and more efficient. From such changes we get new innovations as these new demands for processing power emerge.

In 1998 we saw several developments. The first was SLI (Scan Line Interleave). Two 3Dfx Voodoo² 3D accelerator cards could be hooked together, which allowed two or more cards to work at the same time. Later that year, the way scenes were rendered evolved through object manipulation via transform and lighting engines. This birthed cards such as the GeForce 256.

Soon after that, graphics cards got another boost from the memory manufacturers, as memory intensive graphics cards could tap into the power of DDR. Double Data Rate memory took some time to catch on, but we have had DDR2 and GDDR3 (for graphics only) for some time. This brings us to today; ATI unleashes its latest graphics champion, the Radeon X1950XTX, which uses GDDR4 memory for the first time.

Radeon Gets A Face Lift

Both ATI and Nvidia have slowed their progress in anticipation of Direct3D 10. ATI is working on its R600 processor and Nvidia has its G80, so both should be ready when DX10 parts are needed. With the major halt like a new API, the card makers have shifted to half steps, much like ATI did with its Radeon X850 before the launch of R520, the Radeon X1000 series processors. Nvidia implemented two GeForce 7900 GT processors into a single card with the Geforce 7950 GX2. ATI has followed suit by launching a revised version of its Radeon X1900 graphics card.

Overall, there are only a few major differences. The first thing to catch your eye is the new cooler design. After the advent of loud coolers on the Radeon X850 and X1000 series high performance cards, ATI heard the cry of its constituents and created a heavy copper heat sink with a heat pipe and a much quieter fan. The cooler continues to draw air from inside the case and expel it out the back, but it uses its larger fan's location at the very end of the card to pull air from farther within. It pushes the air through an array of fine fins connected to the heat pipe.

While the new thermal design performs well and looks "cooler," that is the only place where the ATI Radeon X1950 looks much different. Technologically, the processor is the same as the Radeon X1900. It has 48 pixel shader units, which support Shader Model 3.0, and can render HDR and AA at the same time, HD video and it maintains its 384-million transistor count. There were some rumors that the R580+ would be manufactured on an 80-nm process, but the first Radeon X1950 cards to hit the market will continue to utilize a 90-nm process.

RAM Renovations

The other major change is the move to GDDR4 memory. Interestingly, Nvidia was the first company that manufactured a video card equipped with DDR on its GeForce 256 DDR, and DDR3 on the GeForce FX 5700 Ultra. ATI has taken the lead with new memory integration and presents GDDR4 on its Radeon X1950 graphics cards. By utilizing this new memory type, ATI becomes the first graphics company to break the Gigahertz barrier for graphic memories operating at 1 GHz (1,000 MHz, which is effectively 2 GHz DDR).

GDDR4 is built on a smaller fabrication process and will natively consume less energy per clock than GDDR3 memory modules. Smaller process continues to elevate the frequency speed threshold as the electrical gates are smaller and can switch faster. ATI claims that the power savings from the incorporation of new modules will be virtually nullified since the clock speeds are so much faster. Power consumption should therefore remain the same, but the memory bandwidth and performance should increase.

With the other SDRAM specifications, GDDR4 is also an open standard. Many companies participated in the design project, including ATI and Nvidia, as well as memory manufacturers such as Hynix, Infineon, Micron and Samsung. While it is an open standard that companies can choose to produce, there are still a limited number of suppliers at the moment. This should ramp up and we could see the price of GDDR4 start to come down. For this reason, ATI has only implemented this technology into its high-end products.

The advantage of higher clock frequencies means that the amount of data that can flow from the modules into and out from the 512 bit wide ring bus can be expanded. While the external bus to the modules remains at 256 bits wide and module densities remain the same as GDDR3 (512 Mb or 32Mb x 32), the maximum theoretical bandwidth of the GDDR4 memories is now 64 GB/s. That is a 25% bandwidth increase or 13 GB/s over DDR3 at 800 MHz. According to Joe Macri, senior director of circuit technologies at ATI, in the near future these memories could be pushed up to 1.35 GHz, and Samsung has stated that it could be pushed up to 1.5 GHz.

The latest innovations in GDDR come from tweaks such as lower voltage requirements, data bus inversion, widened to a 8-bit data prefetch (versus previous 4-bit prefetch), and a longer fixed burst lengths allows for a better signal in during transmissions and mean power savings. We will continue this discussion in another article but wanted to get the highlights out for the launch.

Meet The New Cards

Radeon X1950 CrossFire Edition

There are many cards that are being announced today. The flagship Radeon X1950XT and the Radeon X1950 CrossFire Edition both have a core clock speed of 650 MHz. This is a change from previous CrossFire Edition card specifications as they would generally be slower than the flagship card. This is good as the cards no longer will need to clock to a comparable speed or in other words, the XTX version will no longer have to go slower anymore.

There are three other cards that ATI is announcing today as well: the Radeon X1900XT 256, the Radeon X1650 Pro and the X1300XT. The first card is the Radeon X1900XT 256, and it is equivalent in all regards to the previous flagship card Radeon X1900XTX except that it has 256 MB of GDDR3 memory instead of 512 MB. While there isn't a CrossFire-compatible card for the Radeon X1900XT 256, there is the strong possibility that it will later be able to double up in CrossFire mode via a software implementation.

The Radeon X1650 Pro is the identical card as the Radeon X1600XT, but with a new name and price. The same goes for the Radeon X1300XT, which formerly was the X1600 Pro. As with many new launches, the processors from previous generations find themselves on new boards with different names and credentials. In this case people are merely getting a price drop with a facelift.

We only received the Radeon X1950XTX, X1950 CrossFire Edition, and X1900XT 256 for testing. We are waiting on the Radeon X1650 Pro and X1300XT. The chart below shows the differences in the entire family. We will provide more coverage once we receive the additional cards.

ATI Radeon X1950, X1900, X1800, & X1600 Specifications
	X1950XTX	X1900XTX	X1900XT	X1900XT 256
Process Technology	90nm	90nm	90nm	90nm
Number of Transistors	384M	384M	384M	384M
Core Clocks (Vertex, Core) in MHz	650/650	650/650	625/625	625/625
Memory Clock (MHz / data rate)	1000/2000	775/1550	725/1450	725/1450
Vertex Shaders (#)	8	8	8	8
Pixel Shaders (#)	48	48	48	48
ROPs (#)	16	16	16	16
Texture units	16	16	16	16
Frame Buffer Size	512 MB	512 MB	512 MB	256 MB
Memory Bandwidth (GB/sec)	64	49.6	46.4	46.4
Vertices/Second (Millions)	1300	1300	1250	1250
Pixel Fill Rate (# ROPs x clk) in Billions/sec	10.4	10.4	7.2	7.2
Texture Fill Rate (# Texture units x clk) in Billions/sec	41.6	41.6	40	40
RAMDACs	400 MHz	400 MHz	400 MHz	400 MHz
Bus Technology	PCI Express	PCI Express	PCI Express	PCI Express
	X1800XT	X1800GTO	X1600XT
Process Technology	90nm	90nm	90nm
Number of Transistors	321M	321M	157M
Core Clocks (Vertex, Core) in MHz	625/625	470/430	440/400
Memory Clock (MHz / data rate)	750/1500	500/1000	690/1380
Vertex Shaders (#)	8	8	8
Pixel Shaders (#)	16	12	12
ROPs (#)	16	16	16
Texture units	16	16	4
Frame Buffer Size	512 MB	256 MB	256 MB
Memory Bandwidth (GB/sec)	48	32	44.16
Vertices/Second (Millions)	1300	940	880
Pixel Fill Rate (# ROPs x clk) in Billions/sec	8.8	6.88	6.4
Texture Fill Rate (# Texture units x clk) in Billions/sec	10	7.52	28.16
RAMDACs	400 MHz	400 MHz	400 MHz
Bus Technology	PCI Express	PCI Express	PCI Express

NVIDIA GeForce 7950, 7900, 7800, & 7600 Specifications
	7950GX2	7900GTX	7900GT	7800GTX 512
Process Technology	90nm	90nm	90nm	110nm
Number of Transistors	556M	278M	278M	302M
Core Clocks (Vertex, Core) in MHz	500/500	700/650	470/450	550/550
Memory Clock (MHz / data rate)	600/1200	800/1600	660/1320	850/1700
Vertex Shaders (#)	16	8	8	8
Pixel Shaders (#)	48	24	24	24
ROPs (#)	32	16	16	16
Memory Interface	256-bit	256-bit	256-bit	256-bit
Frame Buffer Size	1 GB	512 MB	256 MB	512 MB
Memory Bandwidth (GB/sec)	76.8	51.2	42.2	54.4
Vertices/Second (Millions)	2000	1400	940	1100
Pixel Fill Rate (# ROPs x clk) in Billions/sec	16	10.4	7.2	8.8
Texture Fill Rate (# pixel pipes x clk) in Billions/sec	24	15.6	10.8	13.2
RAMDACs	400 MHz	400 MHz	400 MHz	400 MHz
Bus Technology	PCI Express	PCI Express	PCI Express	PCI Express
	7800GTX	7800GT	7800GS AGP	7600GT
Process Technology	110nm	110nm	110nm	90nm
Number of Transistors	302M	302M	302M	178M
Core Clocks (Vertex, Core) in MHz	470/430	440/400	375/375	560/560
Memory Clock (MHz / data rate)	600/1200	500/1000	600/1200	700/1400
Vertex Shaders (#)	8	7	6	5
Pixel Shaders (#)	24	20	16	12
ROPs (#)	16	16	8	8
Memory Interface	256-bit	256-bit	256-bit	128-bit
Frame Buffer Size	256 MB	256 MB	256 MB	256 MB
Memory Bandwidth (GB/sec)	38.4	32	38.4	22.4
Vertices/Second (Millions)	940	770	562.5	700
Pixel Fill Rate (# ROPs x clk) in Billions/sec	6.88	6.4	3	4.48
Texture Fill Rate (# pixel pipes x clk) in Billions/sec	10.32	8	6	6.72
RAMDACs	400 MHz	400 MHz	400 MHz	400 MHz
Bus Technology	PCI Express	PCI Express	AGP 8X	PCI Express

While the technical launch is today, cards will not hit the shelves until September 14. Public relations manager Will Willis stated that "Card orders can be taken on August 23 but they won't ship until September 14." So if you want one you will have to wait but you can place your order today.

Test Setup

As we mentioned before, we did not have the cards for more than a day and a half. While shipping issues hindered our efforts, we wanted to get out as much information into your hands as possible within the short timeframe available. We will have the multi-processor results shortly and in that article we will discus more about power requirements and how some of these new memory changes impact performance and power consumption.

System Hardware
Processor(s)	AMD Athlon 64 FX-602.6 GHz, 1.0 GHz HT-Link, 1 MB L2 cache
Platform	Nvidia : Asus AN832-SLI PremiumNVIDIA nForce4 SLI, BIOS version 1205Nvidia : Asus A8R32-MVP PremiumATI Xpress 3200 CrossFire, BIOS version 0404
RAM	Corsair CMX1024-4400Pro2x 1024 MB @ DDR400 (CL3.0-4-4-8)
Hard Drive	Western Digital Raptor, WD1500ADFD150 GB, 10,000 rpm, 16 MB cache, SATA150
Networking	On-Board nForce4 Gigabit EthernetOnboard Marvell Gigabit Ethernet
Graphics Cards
ATI Based Cards	ATI Radeon X1950XTX 512 MB GDDR4650 MHz Core1,000 MHz Memory (2.00 GHz DDR) ATI Radeon X1950 CrossFire Edition 512 MB GDDR4650 MHz Core1,000 MHz Memory (2.00 GHz DDR) ATI Radeon X1900XTX 512 MB GDDR3650 MHz Core775 MHz Memory (1.55 GHz DDR) ATI Radeon X1950XTX 256 MB GDDR3650 MHz Core775 MHz Memory (1.55 GHz DDR)
Nvidia Based Cards	XFX GeForce 7950GX2 1 GB GDDR3570 MHz Core775 MHz Memory (1.55 GHz DDR) PNY GeForce 7900GTX 512 MB GDDR3675 MHz Core820 MHz Memory (1.64 GHz DDR) XFX GeForce 7900GT 256 MB GDDR3675 MHz Core815 MHz Memory (1.63 GHz DDR) PNY GeForce 7800 GTX 256 MB GDDR3430 MHz Core600 MHz Memory (1.20 GHz DDR) EVGA GeForce 7800 GT 256 MB GDDR3445 MHz Core535 MHz Memory (1.07 GHz DDR) XFX XXX GeForce 7600 GT 256 MB GDDR3590 MHz Core800 MHz Memory (1.60 GHz DDR)
Power Supply	PC Power & Cooling Turbo-Cool 1,000W
System Software & Drivers
OS	Microsoft Windows XP Professional 5.10.2600, Service Pack 2
DirectX Version	9.0c (4.09.0000.0904)
Platform Driver	nForce 6.70
Graphics Driver(s)	ATI - Catalyst Beta (New Cards)ATI - Catalyst 6.6 WHQLNVIDIA - Forceware 91.33 Non-WHQL (7950GX2 Only)NVIDIA - Forceware 84.21 WHQL

Benchmarks Results

3DMark05

In 3DMark we tested at all resolutions with antialiasing and anisotropic filtering disabled in the first run. The second time around we enabled the image quality with 4xAA and 8xAF.

Here we can see that the power of two GeForce 7900 processors configured in the GeForce 7950GX2 is the top dog in 3DMark. We must say that the power of two processors delivers roughly a 15% gain at 1024x768 and 1280x1024 and move up to a 30% gain at the higher resolution. The GeForce 7950GX2 is 20% more expensive and shows some value at the higher resolutions. With filtering turned on the variance between the two top contenders widens with the Nvidia card pulling ahead. Ironically, the ATI Radeon X1900XT 256 does the opposite with its competition the GeForce 7900GT. At the lower resolutions, the Radeon falls slightly behind and then finishes strong at the higher resolutions.

Doom 3

Doom 3 is typically dominated by Nvidia as its core design was matched for this game. Its double Z buffering for stencil shadows has always given it an edge. Both new cards do very well in this title. With antialiasing disabled, the ATI Radeon X1900XT 256 wins until 1600x1200 and beyond. The card outperforms the GeForce 7950GX2 but falls behind the GeForce 7900 cards (GTX and GT). Even at XHD gaming levels, the new ATI cards do very well. With antialiasing turned on to four times sampling it takes second place and the Radeon X1900XT 256 takes fourth.

Half-Life 2: Lost Coast

We ran the usual battery of test but ran into one problem with Half-Life 2: Lost Coast. We utilized a special driver which was laden with the "chuck patch" that allows us to test high dynamic range with antialiasing enabled. On those results you will see zeros for the results as the game would not load to the menu screen properly. ATI has their performance labs trying to replicate the issue. We hope to have a resolution soon. For now the cards get an incomplete until they take the test.

Black And White 2

Where Doom 3 is dominated by Nvidia based cards, Black & White 2 with its heavy use of pixel shaders for hair, foliageand water shows ATI in the driver's seat. The results are no different here. The faster memory yields two to three frames faster than the Radeon X1900XTX and dominates up until the extreme range where the brute force of two graphics processors wins the 2048x1536 and 2560x1600 categories.

F.E.A.R

F.E.A.R. is generally a title that Nvidia-based cards do well in. What is interesting to note in these results is how well the ATI cards did at the higher resolutions. With AA and AF enabled it can go toe to toe with the GeForce 7950GX2 and stays above 30 frames per second at every resolution. Nvidia still holds the F.E.A.R. crown at lower resolutions but from the middle on up, the ATI cards perform well.

Oblivion

We use Elder Scrolls IV: Oblivion as the ultimate torture test. The only settings we don't maximize are HDR and soft shadows. We disable HDR due to the fact that Nvidia cannot render HDR with antialiasing (ATI cannot either without the "chuck patch") and we disable soft shadows because they don't appear correctly as shadows from the back of a character's head can cast a shadow that can be seen on their face. This can make the women appear to have beards.

In our outdoor scene there are long lines of sight, day is changing into night, and there is foliage swaying in the breeze. This has a severe impact on performance. This is where raw horsepower can muscle its way through this test. That is exactly what the GeForce 7950GX2 can do and chew through. While none of the cards are "playable" (meaning that they average 30 frames per second or more) at anything higher that 1024x768, it demonstrates how well the future of games with many moving objects and techniques could be.

Clearly the Nvidia GeForce 7950GX2 is the winner but for $100 more than the Radeon X1950XTX, the cost is not justified. When we switch to the indoors, the ATI cards beat the Nvidia cards and even the Radeon X1900XT 256 gives the GeForce 7950GX2 some blows.

Conclusion

We would have liked to have more time with the cards than we had, but will come back with more reviews. This preliminary look into the single-card operations bodes very well for the red company from up north. At $449 the Radeon X1950XTX is a very nice buy. If you plan an upgrade in the near future, this new addition sweetens the deal.

If you are looking for more for a bit less, the Radeon X1900XT 256 is right up your alley. At a suggested retail price of $279, these look to be a good buy for those who seek a solid gaming card for a less-than-astronomical price.

Time was not a luxury we could afford, so stay tuned for the follow up article with the CrossFire results in comparison to SLI and Quad.