Introduction

Although the average high-end notebook comes equipped with 16GB or 32GB of basic DDR4-2133 memory, gaming-notebook manufacturers are always looking for ways to boost their specs versus the competition's. Nowadays, main system memory is no exception.

In the case of laptop memory, those memory modules are known as SODIMMs, physically shorter in size than their desktop equivalents. The organization that sets DDR4 standards, JEDEC, has gone two steps better than the typical DDR4-2133, with DDR4-2400 and DDR4-2666 both being current SODIMM options, and both operating at the same 1.20V DDR4 standard. Plus, some notebook makers have gone past that, bringing their machines support for gaming-minded memory modules that support XMP memory-overclocking profiles.

We got our hands on some of the latest single SODIMM modules and twin-SODIMM kits to see how much additional performance we could uncover, and whether upgrading to one versus the other (assuming you have the choice to do so) might be worth the differences in price among them. Given the ramp-up in RAM prices at the time we published this, the case for upgrading purely on speed grounds is harder to make than ever. But let's get to quantifying.

The Single DIMMs We Tested

Most notebooks are based on dual-channel memory architecture, but some thin-and-light notebooks have only a single channel of memory filled, along with a single DIMM slot open and available for upgrades. Because running two sticks in dual-channel mode provides more performance, all else being equal, in that upgrade situation most users will get the greatest performance benefit by matching their "additive," second upgrade module to the memory that’s already installed.

As a result, if that describes you, you might be in the market for a single-DIMM memory kit. We snagged three DDR4-2400 and DDR4-2666 kits to do some comparison testing. Let's take a brief look at each of our single-DIMM contenders...

Patriot Viper 2666 1x16GB (PV416G266C8S)

The only 16GB module we received for testing and available near its low MSRP ($170 for 16GB), Patriot’s PV416G266C8S is great for maximizing RAM in laptops with limited slots.

In addition to the obvious capacity benefits versus 8GB, 16GB DIMMs (including SODIMMs) are always organized as "dual-rank," which—though it’s a little complicated to explain in an overview like this—results in improved memory-controller performance.

Patriot adorns these modules with a heat spreader that’s slightly thicker than the foil-backed stickers on typical DIMMs, even though the added thickness could lead to fitment issues in laptop memory-slot arrangements where modules are tightly stacked or overlapping one another. DDR4-2666 CAS 18 timings are available through both standard SPD programming and a 1.20V XMP profile.

Transcend 2400 1x8GB (TS1GSH64V4H)

Transcend’s TS1GSH64V4H is a standard DDR4-2400 SODIMM with typical JEDEC SPD timings of 17-17-17-39 at 1.20V.

This DIMM differs from our other single-module DDR4-2400 sample in that it’s dual-rank, having the requisite eight ICs per side.

Super Talent 2400 1x8GB (F24SA8GM)

Super Talent’s single-rank F24SA8GM has fewer ICs and a lower price than the above Transcend module. That said, it's still rated at the same JEDEC-standard DDR4-2400 data rate, with 17-17-17-39 timings at 1.20V.

This particular kit is sold in single modules only, and Super Talent validates these in fours to ensure multi-channel stability. So, theoretically, you should be able to pair any two or four modules together without issue.

Because we received several of these, we have tested them in both single- and dual-DIMM configurations. We also used them as the baseline memory for our recent DDR4 quad-channel SODIMM kit review.

The Dual-Channel Kits We Tested

Buyers who can afford to replace all their memory at once, or who are ordering from a company that builds completely custom laptop configurations, can get their best performance benefit by installing matching module pairs. In this regard, SODIMM-equipped mini gaming desktops can be treated the same way as gaming notebooks that are based on desktop processors. (In systems like these, support for higher-frequency memory and XMP modes is more common.)

Below, let's run through the dual-DIMM kits we looked at. One thing you'll notice: a profusion of mostly similar G.Skill Ripjaws kits. Why? We tested four sets at differing RAM speeds to get a sense of the ramp-up when going from one speed to another in a given SODIMM family.

G.Skill Ripjaws 3800 16GB (2x8GB, F4-3800C18D-16GRS)

Pulled from the toothsome grip of G.Skill’s same-speed quad-channel kit, the Ripjaws F4-3800C18D-16GRS dual-channel kit came in at an MSRP of $229 (a while back, anyway; more on that in a moment). The DDR4-3800 data rate and 18-18-18-38 timings come from its single 1.35V XMP profile, whereas its standard 1.20V SPD values top out at DDR4-2400.

The thing is, in the time between our tests and our writeup here, this kit has become very hard to find. Our tester did find a kit for sale below MSRP before these mostly vanished from the Internet, so you may have to do some concerted searching before you can land one. At this writing, alas, we couldn't find a set for sale online, but that may change by the time you read this.

G.Skill Ripjaws 3000 16GB (2x8GB, F4-3000C16D-16GRS)

Bearing the same $229 MSRP as its fastest kit (but selling for well above that at this writing), G.Skill’s F4-3000C16D-16GRS supports two DDR4-3000 XMP profiles (16-18-18-43, and 18-18-18-43) to aid compatibility with a variety of XMP-capable notebooks. Three SPD codes are also available, using basic JEDEC timings: 2400, 2133, and 1866.

All configurations use the DDR4 standard of 1.20V. Our testing was done at the DDR4-3000 CAS 16 XMP setting.

HyperX Impact 2666 16GB (2x8GB, HX426S15IB2K2)

Boasting a friendly $194 MSRP, HyperX's Impact model HX426S15IB2K2 also boasts enhanced timings at DDR4-2666 (15-17-17-35), DDR4-2400 (14-16-16-32), and DDR4-2133 (12-14-14-28). Each of those timing sets is two CL cycles quicker than JEDEC-standard, and the highest bandwidth setting is repeated in an XMP table to ease compatibility with XMP-capable systems.

On this kit, all frequency and timing sets use the 1.20V DDR4 standard to improve both compatibility and notebook battery life.

G.Skill Ripjaws 2666 16GB (2x8GB, F4-2666C18D-16GRS)

With an MSRP of $229, G.Skill’s model F4-2666C18D-16GRS runs at JEDEC DDR4-2666 standard 18-18-18-43 timings and 1.20V. With eight DRAM ICs per side, its dual-rank design is its chief advantage over competing modules in this test set.

G.Skill Ripjaws 2400 16GB (2x8GB, F4-2400C16D-16GRS)

A step down the family tree from the G.Skill model above, the company's JEDEC-standard DDR4-2400 model F4-2400C16D-16GRS has an MSRP of $185 and runs at the 1.20V DDR4 industry standard.

MORE: Best Memory

MORE: DDR DRAM FAQs And Troubleshooting Guide

MORE: All Memory Content

How We Tested and Benchmarked DDR4 SODIMM Kits

To run our SODIMM tests, we needed a platform that met three criteria. First, it would allow for easy changes in DIMM configuration; that is, swapping modules in and out quickly. Second, it would support both XMP and SPD timings. And third, it would approximate the highest possible performance available in a notebook.

The need to make easy changes in our DIMM configurations, paired with the need to simulate powerful laptops along the lines of boutique models that take desktop parts, forced us to deploy a desktop motherboard and a matching six-core processor for our testing. To approximate the memory bandwidth of the latest Core i7-8700K-based gaming notebooks that use desktop CPUs, we limited the maximum memory configuration to dual-channel mode. Here's a look at our setup...

Toss in a few bandwidth-affected benchmarks, and we’re ready to go!

Sandra Memory Bandwidth and Cache & Memory Latency benchmarks are great tools to verify that memory is reaching anticipated performance levels regarding bandwidth and latency. Latency is measured in clock cycles, so that both shorter cycles (faster frequencies) and fewer cycles between operations will improve real-time performance. Similarly, because the bandwidth benchmark sends multiple data packets, lower latency settings increase the number of packets per second to improve bandwidth.

The "real-world" Adobe and Microsoft benchmarks mentioned above include timed applications, where less time to complete the task equals more performance, and in the single game (F1 2015), more FPS means more performance. The balance between tests that are greatly affected and less affected by memory performance is meant to represent a wider range of user experience.

MORE: Best Memory

MORE: DDR DRAM FAQs And Troubleshooting Guide

MORE: All Memory Content

Single-DIMM SODIMM Laptop Memory Kits - Tests and Benchmarks

A note on the performance-result tables to follow: We organized the tables by highest total capacity, followed by highest data rate.

A comparison of the Transcend and Super Talent modules shows that, when scoring overall value (the "Performance Per Capacity Per Dollar" table), a lower price has a bigger impact than any meaningful improvement in performance does.

Digging a little deeper into our Overall Performance chart, we find that games are the best place to see the dual-rank Transcend DDR4-2400 outpacing the single-rank Super Talent DDR4-2400, and that the Patriot Viper's combination of dual-rank design and higher data rate is best.

Our timed workloads (Adobe CC, Microsoft Office) were barely affected by our test kits' variances in data rate or module design, apart from the fact that PCMark's heavy workload for Photoshop is optimized for more than 8GB of DRAM. Sandra's bandwidth-dependent Cryptography bench, meanwhile, is a near-perfect reflection of bandwidth in these single-channel configurations. You can see the clear stepping from 2400 to 2600 modules.

Following from these tests are our individual assessments of the three single-DIMM kits we tested. In short, the Patriot kit is a fine performance pick if all you have to work with is a single SODIMM slot, and your target laptop can handle the timings; the Super Talent makes sense if all that you're after is capacity.

MORE: Best Memory

MORE: DDR DRAM FAQs And Troubleshooting Guide

MORE: All Memory Content

Dual-Channel SODIMM Laptop Memory Kits - Tests and Benchmarks

The fastest memory in our test, the Ripjaws DDR4-3800 in its 1.35V XMP configuration, barely outpaced its stock-voltage DDR4-2666 sibling in overall performance. Digging deeper, we find that our game testing illustrated the biggest performance variances, and that the Ripjaws DDR4-2666 won that round. The major differentiating factor was in the number of DRAM ICs, illustrated best by G.Skill's DDR4-2666 kit: It uses 16 ICs to fill two ranks of DRAM per SODIMM.

Digging still deeper into the charts, we find that the “Photoshop Heavy” benchmark in our Adobe CC test suite makes the strongest case for using DDR4-3800 in real-world applications. This 3800 G.Skill set, the fastest module kit in our test lot, also has the least latency, because latency is set in cycles (and faster cycles take fewer nanoseconds). The extra bandwidth available from high data rates also helped the fastest modules cruise through Sandra Cryptography, and Sandra Memory Bandwidth shows us the peak available performance that most real-world applications are incapable of extracting.

Don't forget, however, that games are the tests in which fast memory matters most, and that G.Skill's dual-rank DDR4-2666 won that round without the need for XMP. So, to reiterate: The presence of dual-rank trumped other concerns in that test.

Here's a look at the five individual kits we tested...

MORE: Best Memory

MORE: DDR DRAM FAQs And Troubleshooting Guide

MORE: All Memory Content

What to Consider Before Updating Your Laptop Memory

On the preceding pages, we walked you through our benchmark tests for eight SODIMM kits and an overview of each product, starting with the single-DIMM results and followed by the dual-DIMM ones. You can dig into the details there on a product-by-product basis, but here's our bottom line: Buying SODIMMs is way more about what you can install than what you might want to install. As a result, you generally shouldn't obsess over maximizing SODIMM speed, except under certain very narrow circumstances.

The money you might spend maximizing RAM speed isn't likely to make a huge real-world difference. That's generally true in the desktop realm, as well, except when playing PC games on integrated graphics, as we saw in our AMD Raven Ridge memory scaling story, AMD Ryzen 5 2400G Tests: How Much Does Memory Impact Gaming? Given that most gaming-focused laptop users will almost certainly have a dedicated graphics chip, maximum memory speed isn’t going to make a major difference.

In the end, once you’ve whittled down the memory modules that will work with your system, if you’re looking for the best balance of performance and price, our advice is (obviously) to make sure you’re running at least two SODIMM sticks (if at all possible) for the extra bandwidth that comes from doubling your memory channels. Then, to nudge the performance needle a little further forward, consider dense dual-rank kits with lots of integrated circuits, to increase the bandwidth even more.

Beyond that, if you want to splurge on high-speed, overclockable memory and your system supports it, you certainly have options. But given the high price of memory these days, and the diminishing returns of expensive RAM with all the possible megahertz, we’d suggest saving that money for other upgrades (say a faster/roomier boot drive) or saving your cash for a future laptop or SFF desktop. Perhaps, by the time you have enough for a new machine, memory prices will have fallen, and new and better graphics chips will be available. We can only hope, but hoping for a better hardware tomorrow might just be a better bet than overspending on expensive RAM today that doesn’t give you much (if any) noticeable benefit.

OK, But I Still Want to Upgrade. Which of These Kits Are Best?

Of the eight kits tested here, to a certain extent the "best" kits are the ones you can install, or whose timings your laptop supports.

Also, let's take the most common upgrade situations.

Some folks are limited by the vagaries of an existing installed module. They're just trying to match the one they have installed as best they can to boost RAM capacity. Their "value" calculation is limited by what they already have installed. They may have to match their existing memory since the installed module is not easily accessible (i.e., they can't swap it out), or their budget precludes swapping out the existing memory; they simply need to augment, not replace.

Others have just one SODIMM slot to play with, and are aiming to swap out an existing lesser-capacity module in that single slot with a bigger-capacity one. Of course, they are working within the spec limitations of the laptop, and that limits their choice pool. Their sole objective is to buy a bigger module than what they had installed before (and maybe one that is faster, too, if their system supports it).

Other upgraders may have two slots to play with, and mean to fill both slots with new modules. That's a similar objective to the previous upgrader's, but again limited by what the system supports.

So, What To Do?

To make sense of these disparate upgrade and/or install situations, we have to split up our recommendations.

If you are in a situation where you have the flexibility to install high-speed RAM, the subset to consider is way different. Assuming you can use it, we like the 2666 Ripjaws kit for its balance, as well as the HyperX...

That said, most of the modules tested here offer solid value to laptop users who simply need more memory capacity. And the single modules could be particularly useful in those upgrade situations where you have only a single slot to work with. For these users, Super Talent DDR4-2400’s lower price--assuming it stays low, relative to the rest--will be a big draw.

That said, none of these modules raised a red flag on performance. And indeed, the sheer benefit of more RAM capacity than you had before should outweigh the mild performance differences among specific apps, assuming you choose a suitable kit within the limitations of your laptop's memory support.

MORE: Best Memory

MORE: DDR DRAM FAQs And Troubleshooting Guide

MORE: All Memory Content

Best Power Supply Units For CryptoCurrency Mining At Home

Caution

We advise against constantly operating your PSU at maximum capacity, especially under high ambient temperatures. Besides lower efficiency, the unit will also suffer a dramatically reduced lifetime.

Without a doubt, the component that takes the most stress in a cryptocurrency mining rig is its power supply. Of course, when you choose to install multiple GPUs for mining, you have to complement them with a powerful PSU or use more than one power supply to deal with the load. Given lots of demand right now, you might find the highest-capacity models out of stock. While it might be tempting to use a lower-capacity PSU pushed closer to its limit, that can lead to catastrophic results, especially since mining rigs often operate unattended. It goes without saying that mining requires a high-quality and ultra-reliable power source. Capacity and PCIe connector count shouldn't be the only factors that influence your purchasing decision. So, to help you pick the best PSU for your mining rig, we're digging deep into our comprehensive database of benchmark results to identify the top models.

A mining PC's PSU needs to have some specific features and specifications if it's to survive the job you're giving it. We're going to use the requirements set forth by Cybenetics in its Mining-Ready PSU project as a foundation for our project.

We define home miners as folks dabbling in cryptocurrency mining on the side. They don't want to spend a fortune building dedicated mining machines. Instead, they're looking for something that'll yield some profit and won't cost much. The home miner's rigs are running somewhere inside the house, subjected to reasonable ambient temperatures. Still, noise can't be allowed to get out of control.

Below you'll find our selections for the best PSUs for cryptocurrency mining, a list of our requirements and recommendations, details on the selected PSUs, and some benchmark numbers based on critical criteria. This page is for those mining at home; the following page is for professional mining.

MORE: Best Deals

Best PSUs For CryptoCurrency Mining At Home

3/9/2018 Update: The Enermax MaxTytan 1250W and EVGA 1600 T2 PSUs have been added.

2/6/2018 Update: We made some changes to the requirements and the recommendations. Corsair AX1600i and Seasonic Prime Ultra Platinum PSUs have been added.

MORE: Best Power Supplies

MORE: How We Test Power Supplies

MORE: All Power Supplies Content

These are the requirements and recommendations we're defining for home-based cryptocurrency mining PSUs:

Requirements

• 1kW or higher capacity.
• 115V: ETA-A- (85-88%) efficiency or higher.
• 230V: ETA-A- 230V (87-90%) efficiency or higher.
• 80 PLUS Gold efficiency or higher.
• If the PSU is certified by Cybenetics, it must have a LAMBDA-S++ (30-35 dB[A]) noise rating, at least. If it isn’t certified by Cybenetics, we will accept all entries that satisfy the other requirements on one condition: if the PSU is certified by Cybenetics in the future and doesn’t achieve a LAMBDA-B or better rating, it will be immediately removed from the list. 
• Lower than 50mV ripple at +12V under full load at increased operating temperatures (>40°C).
• Quality fan (FDB or similar; ideally it should use ball or magnetic bearings).
• At least six 6+2 pin PCIe connectors on PSUs with up to 1.1kW capacity. For PSUs with over 1.1 kW capacity the number of minimum 6+2 pin PCIe connectors is eight and for PSUs with over 1.45 kW capacity the minimum is ten 6+2 pin PCIe connectors.
  
• At least four 4-pin Molex connectors on two at least cables, in 1 kW and stronger PSUs, and six 4-pin Molex connectors in three or more cables for 1.45 kW and higher wattage PSUs. In case the PSU has more 6+2 pin PCIe connectors than the required, we are covered with four 4-pin Molex connectors regardless capacity.
• All peripheral cables should use 18AWG wires minimum.
• For 1.4kW and stronger PSUs, a C19 coupler is required. An AC power cord with 14AWG wires should be used. For lower-capacity PSUs, an AC power cord with at least 16AWG wires is required.
• Support the essential protection features (SCP, OPP), including over-temperature protection.
• Over 17ms hold-up time and an accurate power-good signal, which has to be at least 16ms. The power-good signal has to have at least a 1ms delay, dropping at least 1ms before the rails go out of spec.
• Complete EMI filtering stage (minimum components: 4x Y caps, 2x X caps, two CM chokes, an MOV), along with inrush current protection (an NTC thermistor is required, which ideally should be supported by a bypass relay).
• Impeccable build quality, including quality MOSFETs and high-quality bulk/filtering capacitors (105°C rating and a majority of filtering caps on the secondary side must have >4000h lifetime). The use of polymer caps on the secondary side is preferred.

Recommendations

1. Two EPS connectors
2. Dedicated PCIe cables are preferred, along with 16-gauge wires.
3. The peripheral connectors should have 15cm distance between them.
4. Ideally, every PSU with 1kW or more capacity should use a 14-gauge power cord along with a C19 coupler.
5. One year minimum warranty for 24/7 operation under full load at 30-35°C

In total, we have 15 requirements and five recommendations for PSUs used in home-based cryptocurrency mining PCs. The most important are the efficiency and build quality factors, along with capacity. Inside of a home, where this type of mining rig will operate, climate control maintains comfortable temperatures, so in most cases we accept FDB fans and their derivatives. These don't have reliability issues below 35°C ambient. The fan does have to use a true FDB or high-quality rifle bearing, and not just a plain sleeve bearing. Those are the cheapest and most unreliable solutions for PSUs running 24/7.

Another recommendation is to use your PSU with 230V input whenever possible. Besides 1-2%-higher efficiency, this also reduces the amperage passing through the AC power cord by half. As a result, the cord endures much less stress. If you live in a region with 115V mains and can afford the installation of 230V sockets in your home, definitely use them for your mining ventures.

Best PSUs For CryptoCurrency Mining At Home

PSUs That Didn't Make It (And Why)

The PSUs from the database that failed to satisfy the requirements set forth for cryptocurrency mining at home are listed below.

• be quiet! Dark Pro 1200W: hold-up time
• Cooler Master MasterWatt Maker 1200: power-ok signal
• FSP Aurum PT 1000FM: hold-up time
• Thermaltake Toughpower DPS G RGB 1500W: hold-up time, power-ok signal
• Thermaltake Toughpower Grand RGB 1200W Platinum: power-ok signal
• Enermax Platimax D.F. 1200W: 6x PCIe connectors, power-ok signal (only 0.3ms less than the required though)

Their problem with the majority of failed PSUs has to do with the power-good signals that we measured, which are either lower than the ATX spec recommends or inaccurate (or both). Even in PSUs used for normal tasks, an inaccurate signal can prove fatal. So imagine what happens to your pricey graphics cards if the rails go out of spec and the PSU pretends everything is still good, rather than telling your motherboard to shut down immediately. Very low voltage levels will push the graphics card's and motherboard's DC-DC converters to their limits, potentially frying components. It's a shame to see such high-end PSUs reporting fake power-good signals. Although they satisfy the rest of our requirements, we strongly advise against them.

With a longer hold-up time and an accurate power-good signal, the 1.5kW Toughpower would be ideal for a home mining system since it offers lots of capacity and 10 PCIe connectors on dedicated cables. If Thermaltake listens to our recommendations and fixes this unit, we will surely add it to our list. Cooler Master's MasterWatt doesn't fall short by much. Should the company rectify its shortcomings, it'll only have to tune its power-good signal accordingly. The Enermax Platimax D.F 1200W is only a hair away from the power-ok signal's requirement, however it only has six PCIe connectors which are too few for a PSU of such high capacity.

Disclaimer: Aris Mpitziopoulos is Tom's Hardware's PSU reviewer. He is also the Chief Testing Engineer of Cybenetics, and developed the Cybenetics certification methodologies apart from his role on Tom's Hardware. Neither Tom's Hardware nor its parent company, Purch Media, are financially involved with Cybenetics. Aris does not perform the actual certifications for Cybenetics.

Best Power Supply Units For Professional CryptoCurrency Mining

Pro miners are more extreme users with systems installed in dedicated spaces (farms), many of which lack air conditioning. We have to anticipate higher ambient temperatures as a result (as high as 40ºC in some cases!). Even if a farm is equipped with HVAC, imagine how hard it is to maintain comfortable temperatures, given the massive thermal output of 10s or 100s of fully utilized GPUs. This is why, for our professional mining category, we only accept PSUs equipped with ball-bearing fans. Any FDB-derivative fan is bound to fail quickly because its lubricant evaporates under such harsh conditions. Since mining farms are typically isolated from where people eat, sleep, or work, noise isn't as much of a concern. In fact, we embrace aggressive fan profiles; they're often what keeps a PSU alive under full load in a hot room.

MORE: Best Deals

Best PSUs For Professional CryptoCurrency Mining

3/9/2018 Update: The Enermax MaxTytan 1250W and EVGA 1600 T2 have been added.

2/6/2018 Update: We made some changes to the requirements and the recommendations. The 17ms hold-up time is now a recommendation and not a requirement, allowing the EVGA 1600 P2 to be included in the list.

MORE: Best Power Supplies

MORE: How We Test Power Supplies

MORE: All Power Supplies Content

The more stringent requirements tied to our professional classification limits the number of models that qualify for this top tier. Ball bearing-equipped fans and higher capacity levels are simply must-haves.

These are the requirements and recommendations we're defining for professional-oriented cryptocurrency mining PSUs:

Requirements

• 1.2kW or higher capacity.
• 115V: ETA-A (88-91%) and higher efficiency certification.
• 230V: ETA-A 230V (90-93%).
• 80 PLUS Platinum or higher efficiency.
• Lower than 50mV ripple at +12V under full load at increased operating temperatures (>45°C).
• Double-ball or magnetic bearing fans. All other types are excluded.
• At least eight 6+2 pin PCIe connectors, or ten 6+2 pin connectors on PSUs with over 1.45kW capacity.
• At least four 4-pin Molex connectors on two cables, or six 4-pin Molex connectors on three or more cables for PSUs with over 1.45kW capacity. In case the PSU has more 6+2 pin PCIe connectors than the required, we are covered with 4-pin Molex connectors regardless capacity.
• All peripheral cables should use 18-gauge wires minimum.
• For 1.4kW and stronger PSUs a C19 coupler is required, while an AC power cord with 14AWG wires should be used. For lower capacity PSUs an AC power cord with 16AWG gauges at least is required. The AC power cord requirement applies only if the PSU is provided with an AC power cord.
• Support the essential protection features (SCP, OPP), including over-temperature protection.
• Accurate power-good signal. The power-good signal has to have at least a 1ms delay, dropping at least 1ms before the rails go out of spec.
• Complete EMI filtering stage (minimum components: 4x Y caps, 2x X caps, two CM chokes, an MOV), along with inrush current protection (an NTC thermistor is required, which ideally should be supported by a bypass relay).
• Impeccable build quality, including quality MOSFETs and high-quality bulk/filtering capacitors (105°C rating and a majority of filtering caps on the secondary side must have >5000h lifetime). The use of polymer caps on the secondary side is preferred.

Recommendations

1. Ideally the peripheral connectors should have 15cm distance between them.
2. Dedicated cables with 6+2 pin PCIe connectors are preferred. In case a cable hosts two PCIe connectors, it needs to have 16AWG gauges up to the first connector for the six out of the eight pins at least. 
3. Every PSU with 1.3kW or more capacity should use an AC power cord with 14AWG gauges and a C19 coupler
4. Over 17ms hold-up time and an accurate power-good signal, which has to be at least 16ms. The power-good signal has to have at least a 1ms delay, dropping at least 1ms before the rails go out of spec.
5. One year minimum warranty for 24/7 operation under full load at 35-40°C

The major differences between our home and professional cryptocurrency mining PSU requirements are the capacity, efficiency, and fan specifications. Also, notice that there is no output noise ceiling in this category, and we're also recommending at least one year of warranty coverage for 24/7 operation under full load at 35-40°C. That's going to be difficult to satisfy; we don't know if any manufacturers will be brave enough to guarantee their hardware for continuous operation under full load at 40°C. Usually, only server-class PSUs can withstand those conditions. Finally, the over 17ms hold-up time rule is a recommendation for pro mining PSUs, since hold-up time doesn't play so much in mining applications. However we did keep the accurate power-good signal requirement since it affects the system's overall health.

It goes without saying that, if you own a mining farm with lots of systems, you'll save a ton of money on electricity using 230V mains. In addition, thanks to higher achievable efficiency, 230V input facilitates lower thermal loads, extending the life of your PSUs.

Best PSUs For Professional CryptoCurrency Mining

PSUs That Didn't Make It (And Why)

The list of PSUs that don't qualify for our list includes the following models so far:

• be quiet! Dark Pro 1200W: HDB fan
• Cooler Master MasterWatt Maker 1200: power-ok signal
• Corsair AX1600i: FDB fan
• Enermax Platimax D.F. 1200W: 6x PCIe connectors
• Rosewill Tokamak 1500W: power-ok signal, Ripple
• Silverstone Strider Gold S 1500W: 8x PCIe connectors
• Silverstone Strider Titanium ST1500-Ti: power-ok signal
• SilverStone Strider Platinum ST1200-PT: power-ok signal
• Thermaltake Toughpower DPS G RGB 1500W: HDB fan, power-ok signal
• Thermaltake Toughpower Grand RGB 1200W Platinum: power-ok signal

Most of them suffer inaccurate power-good signals. This is an issue that seems to plague high-capacity PSUs in general. There is a unit (Rosewill Tokamak 1500W) also which demonstrates high ripple at +12V under full load. In order to achieve the highest possible efficiency, some manufacturers use fewer filtering components on the secondary side, and this inevitably leads to lousy ripple suppression. It'd be better if those companies used a more advanced platform instead of cheap tricks.

The SilverStone Strider Titanium with 1.5kW capacity which used a DBB fan, which is ideal for mining purposes, also has an inaccurate power ok signal so it is excluded from the list.

The Corsair AX1600i would be excellent if it used a DBB or magnetic bearing fan; the same goes for the be quiet! Dark Pro 1200W which offers plenty of connectors.

Disclaimer: Aris Mpitziopoulos is Tom's Hardware's PSU reviewer. He is also the Chief Testing Engineer of Cybenetics, and developed the Cybenetics certification methodologies apart from his role on Tom's Hardware. Neither Tom's Hardware nor its parent company, Purch Media, are financially involved with Cybenetics. Aris does not perform the actual certifications for Cybenetics.

This roundup of Intel Z370 Coffee Lake motherboards pulls from our in-depth reviews, and includes recommendations for gaming, value & overall performance.

Intel's Z370 was designed to support Coffee Lake processors, which have more cores than their Kaby Lake predecessors. Z370's featureset is essentially identical to Z270, and both chipsets are designed for the LGA1151 socket, but Z370 utilizes an updated power design to ensure stability for the more powerful CPUs. This change in power regulation, however, prevents Skylake and Kaby Lake processors from working on Z370 motherboards, despite sharing the same socket design.

MORE: Best Deals

Intel Z370 Motherboards

11/30/2017 Update: original publication; we'll be updating this article every time we review an Intel Z370 motherboard.

MORE: Best Motherboards

MORE: How To Choose A Motherboard

MORE: All Motherboards Content

Because Coffee Lake processors have more cores and require more power, the need for a strong power delivery system has also proportionally increased. Our reviews of some Z370 motherboards have shown temps on voltage regulators are significantly higher than on Z270 motherboards. If a board lacks proper cooling, it could cause thermal throttling on high-end Coffee Lake processors.

Memory support will also be a key spec to watch on Z370 motherboards. Z370's memory support is essentially the same as its predecessor, but with six cores and 12 threads, the amount of bandwidth available to each core has been reduced. The additional cores more than make up for any negative impacts this may have, but selecting a motherboard that supports DDR4 at high frequencies will help to ensure your system works at peak efficiency.

EATX Intel Z370 Motherboards

MORE: AMD X370 Motherboard Roundup

ATX Intel Z370 Motherboards

MORE: Intel X299 Motherboard Roundup

MORE: Intel Z270 Motherboard Roundup

Nvidia’s GeForce GTX 1080 Ti is, for all intents and purposes, the company’s fastest desktop gaming graphics card - Titan Xp, the company's only faster (niche) card, costs ~60% more. The 1080 Ti employs a 12-billion-transistor GP102 processor sporting 3584 CUDA cores, 224 texture units, and 88 ROPs. A base clock rate of 1480 MHz typically accelerates to 1582 MHz under most gaming loads thanks to GPU Boost. And Nvidia’s partners tune those frequencies to differentiate their boards. An aggregate 352-bit memory bus hosts 11GB of GDDR5X memory able to move up to 484 GB/s of information.

In comparison, the vanilla GeForce GTX 1080 utilizes a much less complex GP104 GPU armed with 2560 CUDA cores, 160 texture units, and 64 ROPs. Its 8GB of GDDR5 memory is still high-end by today’s standards. However, you’re looking at the difference between smooth performance at 3840x2160 from GeForce GTX 1080 Ti and a card better suited to 1440p at maxed-out details in the vanilla 1080. You’ll pay about $200 more for the GeForce GTX 1080 Ti, but in a world now unfriendly to multi-GPU configurations, we recommend buying the fastest single-GPU solution you can afford.

How about the competition from AMD? Performance-wise, there is none. Radeon RX Vega 64 is more comparable to GeForce GTX 1080. At least for now, GeForce GTX 1080 Ti is in a class unto itself.

So, if you’re looking for the best graphics performance available, let’s explore some of the available options and why you might consider one GTX 1080 Ti over another.

MORE: Best Deals

Nvidia GeForce GTX 1080 Ti Graphics Cards

10/23/2017 Update: added the MSI GTX 1080 Ti Lightning Z (no award).

MORE: Best Graphics Cards

MORE: All Graphics Content

Available GeForce GTX 1080 Ti cards span a ~$150 range. So why would you consider spending extra when the least-expensive models are already brutally fast? The reasons are loosely interrelated: clock rates, cooling, advanced features, and aesthetics.

You see, Nvidia’s reference specification calls for a base GPU frequency of 1480 MHz. But GP102 can run faster. So add-in board partners experiment with more aggressive settings. A 5% or 6% increase means higher frame rates. And when you’re already looking at top-of-the-line, extra performance isn’t cheap.

Overclocked GPUs generate more heat, creating stability and reliability issues. As a result, the pursuit of performance goes hand-in-hand with upgraded cooling. Some manufacturers turn to triple-slot heat sinks and fans, while others integrated liquid cooling. Along the way, they strive for improved acoustics using larger, slower-spinning fans or beefy radiators.

All of the surface area introduced by these substantial thermal solutions becomes an ideal canvas for decoration. Enter the elaborate fan shrouds, the configurable RGB lighting, and the LCD read-outs.

In the end, you’re left to decide whether Nvidia’s reference specs are good enough, or if you’re willing to pay for a bit more guaranteed performance, the cooling to keep an overclocked config stable, and maybe a little eye candy. Our reviews dig deep into the strengths and weaknesses of each individual implementation, allowing you to choose the attributes that matter most in a GeForce GTX 1080 Ti.

The final analysis of each board is listed below for quick and easy reference (top-awarded cards appear first, and graphics cards with the same award are listed alphabetically):

MORE: Desktop GPU Performance Hierarchy Table

MORE: The History Of Nvidia GPUs

MORE: Nvidia GeForce GTX 1080 Roundup

MORE: Best PC Builds

MORE: Best Cases

MORE: Best CPU Cooling

MORE: Best CPUs

MORE: Best Gaming Desktops

MORE: Best Gaming Laptops

MORE: Best Memory

MORE: Best Gaming Monitors

MORE: Best Motherboards

MORE: Best Power Supplies

MORE: Best SSDs

MORE: Best Virtual Reality Headsets

With the launch of Coffee Lake, Intel's Kaby Lake products have started the slow march towards exiting the market. But there is still value in the aging platform, and the 200-series chipsets and the associated Kaby Lake processors still offer a lot of performance for your everyday needs.

Intel's Z270 chipsets is the most feature rich solution for Kaby Lake processors. It is nearly identical in every detail to the newer Z370 chipset, but it does not support the newer Coffee Lake CPUs, despite sharing the same LGA 1151 socket. Compared to the other 200-series chispets, Z270 has more HSIO lanes, PCIe lanes, and USB 3.0 ports, and it is the only 200-series chipset that supports overclocking CPUs. The Z270 chipset is also the only 200-series chipset that can divide the CPU's PCIe 3.0 lanes between multiple graphics cards.

MORE: Best Deals

Intel Z270 Motherboards

MORE: Best Motherboards

MORE: How To Choose A Motherboard

MORE: All Motherboards Content

Motherboards have dozens of specs that may vary from one board to the next, but there are four key variables that outweigh the rest: CPU, chipset, memory, and power delivery system. Now, depending on your own specific uses, other factors such as the USB support or networking options may be prioritized, but the importance of these other features will vary considerably from one person to the next, whereas everyone should carefully consider the four key variables.

Typically, you first decide if you want to build your new PC around an AMD or an Intel CPU. Next, you'll need to pick a chipset based on its feature set and price. Because this article is focused on Z270 motherboards, we will proceed under the assumption that you've selected an unlocked Intel Kaby Lake processor and the Z270 chipset. If you don't plan to overclock, you may want to consider a lower-end chipset and CPU.

You should also consider the system's memory support. Unless you're building a compact PC, you want to have four DIMM slots to support higher quantities of memory. The maximum memory frequency is also important if you plan to use high-performance DDR4. Officially, Kaby Lake only supports DDR4 at speeds up to 2400MHz, but a Z270 motherboard may support DDR4 DIMMs operating at speeds of 4233MHz or higher.

When selecting a Z270 motherboard, it's crucial to pay attention to the power delivery system, especially if you plan to overclock. An insufficient number of power phases or inadequate cooling can limit your overall system performance and cause the CPU to throttle. In general, a greater number of power phases and larger heatsinks over the MOFSETs, the better. While eight phases is typically sufficient for most, boards with more may produce better overclocking results.

Extended ATX Intel Z270 Motherboards

ATX Intel Z270 Motherboards

Mini-ITX Intel Z270 Motherboards

Introduction & Overview

Nvidia's GeForce GTX 1080 is the company's gaming flagship right now (check out our Nvidia GeForce GTX 1080 Graphics Card Roundup for in-depth analysis of the top models). But its GeForce GTX 1070 is undoubtedly a better value play. After all, the GP104-powered card outperforms the old GeForce GTX 980 Ti for hundreds of dollars less.

There's a lot of variation out there, though. You'll find 1070s selling anywhere from under $400 (£300) to over $500 (£400), and it's amazing to see what Nvidia's add-in board partners have done with this platform. To help you size up the current field, our team continues testing GeForce GTX 1070 cards. For now, we have four different models to compare. And we go deep. Really, each page of this piece could be its own complete review. We focus on the manufacturing quality and technical features of each card, along with power consumption, clock rates, cooling, and acoustics.

The gaming performance of every factory-overclocked board within a given chip class is usually pretty similar to begin with. But that's more true now than ever. This is a result of features like GPU Boost 3.0, which allows manufacturers to safely extract as much headroom as possible from a processor. Very little is left on the table, even if you have access to extreme overclocking hardware.

We will continue to update this roundup as new test samples become available. Some are already in the lab; others are en route (and there's at least one model we need to bring over from Tom's Hardware DE still).

MORE: Best Graphics Cards

MORE: Desktop GPU Performance Hierarchy Table

MORE: All Graphics Content

The cards are presented in alphabetical order, and their order has nothing to do with performance in our roundup or a preference for individual products.

MORE: Best Deals

Benchmark Results

Gaming

We test every card after a suitable warm-up period to avoid unfair differences in GPU Boost frequencies. All benchmarks are run six times; the first one is used to get the GPU hot again.

These cards are all press samples operating at the same settings as retail models in our best effort to ensure one vendor doesn't get a leg up on another using non-representative clock rates.

The following galleries each contain four images, covering two tested resolutions. We put our focus on QHD (2560x1440) and UHD (3840x2160), plotting out average and minimum frame rates for each resolution in separate graphs.

Ashes of the Singularity

Battlefield 4

The Division

Grand Theft Auto V

Hitman

Project CARS

Rise of the Tomb Raider

The Witcher 3

All of the factory-overclocked cards offer similar performance, more or less. That's why our primary focus centers on evaluating the more technical aspects of each board design, along with their coolers. This is where differences in noise, power, and temperatures are most likely to come from.

Power Consumption

We begin by comparing the power consumption of each card in our gaming loop, stress test, and at idle. Depending on the power targets specified by each manufacturer, we sometimes measured substantial differences. We're ignoring the decimal places in our bar graphs, since those values are too small and would be within measurement tolerances.

We also found that some cards with lower power targets started to throttle during our stress test, resulting in lower power consumption numbers. This did not, however, have a negative impact on general gaming performance for any of the tested cards, as the stress test merely represents a worst-case scenario.

We confirmed that MSI's retail cards will ship with a slightly lower power target (max. 240 to 250 watts) after an internal discussion and evaluation of our measurements.

This also applies to the BIOS versions with OC mode enabled by default, which employ a roughly 20 MHz-higher base and GPU Boost frequency. In the interest of fairness, we tested both MSI cards using normal mode, without the overclocked base and GPU Boost rates. This doesn't affect our power consumption measurements, though.

Noise

For the following comparison, we divide all of the gallery's bar graphs into gaming loop and idle, even if "noise" in practice spans a wide spectrum and the "character" of the sound varies a lot, with each card having highly individual results. Therefore, it is important not to compare just the absolute numbers, but also the frequency spectrum we're presenting.

Many of the cards implement a semi-passive mode, where their fans remain off when the card is idle. Thus, we refrained from taking measurements in that state. Even in our anechoic chamber, levels of 22 dB(A) and below merely represent ambient noise.

Temperatures

For this comparison, we divide the gallery's bar graphs into gaming loop, stress test, and peak temperatures measured on the MOSFETs.

We occasionally compared the temperatures on our benchmark table with those measured inside a closed case and found them to be no more than two or three Kelvin higher.

Since temperatures in a closed case also depend heavily on the enclosure's cooling performance, the only representative and reproducible values are those measured on our benchmark table. Those are the ones we compare.

MORE: Best Graphics Cards

MORE: Desktop GPU Performance Hierarchy Table

MORE: All Graphics Content

Nvidia GTX 1070 Founders Edition

The GeForce GTX 1070 Founders Edition is based on the same GPU as Nvidia's GeForce GTX 1080, only with 25% of its CUDA cores and texture units disabled. As we learned in our Nvidia GeForce GTX 1070 8GB Pascal Review, though, the processor's back-end still serves up 64 ROPs and 2MB of L2 cache. Also, GDDR5X memory is replaced with standard GDDR5, while the 1070 sports a less sophisticated cooling solution and a lower power target of 150W.

The Founders Edition board is actually what Nvidia used to call its reference design. This time around, however, the company decided it wanted a piece of the premium card sales as well. Renaming its own offering is supposed to help sidestep the negative connotation that sometimes accompanies reference implementations.

Technical Specifications

MORE: Best Graphics Cards

MORE: Desktop GPU Performance Hierarchy Table

MORE: All Graphics Content

Exterior & Interfaces

The injection-molded aluminum cooler cover is metallic silver and black in color. Its design exudes quality, but the metal construction also contributes to the card's relatively high weight of more than 35 ounces (one kilogram).

The back of the board is covered by a two-piece plate. It's purely aesthetic and does nothing for cooling. If you have a couple of cards next to each other in SLI or if you'd simply prefer more airflow, the backplate can be unscrewed without a problem.

There's an illuminated GeForce GTX label on top of the card, and an eight-pin auxiliary power connector is positioned toward the end.

Nvidia's angular and rugged design is certainly a matter of taste. But it certainly stands out. In a windowed case, you can count on turning heads.

Peeking down through the card's end reveals the heat sink's fins. The mounting frame offers up to three screws for attaching additional brackets. You'd use those for stabilization in a case.

The slot plate features five outputs, four of which can be used simultaneously in a multi-monitor setup. In addition to one dual-link DVI-D interface, the rear bracket also exposes an HDMI 2.0 output and three DisplayPort 1.4-compatible connectors. The rest of the plate is peppered with openings to encourage better airflow.

Board & Components

In addition to the one phase for memory, only four out of six available power phases for the GPU are implemented.

Nvidia leans on Samsung for its GDDR5 memory. Eight K4G80325FB-HC25 modules are each able to store up to 8Gb (32x 256Mb). Voltages are specified between 1.305V and 1.597V, depending on the clock rate.

The 4+1-phase system relies on the sparsely documented µP9511P PWM controller. Since this controller can't communicate directly with the VRM's phases, Nvidia utilizes 53603A chips for solid PWM drivers (gate drivers) and controlling the power MOSFETs (primarily of type 4C85N).

Two capacitors are installed right below the GPU to absorb and equalize peaks in voltage. The board design looks tidy and well thought-out.

Power Results

Before we look at power consumption, we should talk about the correlation between GPU Boost frequency and core voltage, which are so similar that we decided to put their graphs one on top of the other. This also shows that both curves drop as the GPU's temperature rises.

After warm-up, the GPU Boost frequency flirts with 1.7 GHz under load. This is mirrored by our voltage readings. While we measured up to 1.031V in the beginning, that value later dropped as low as 0.812V, if only temporarily.

Combining the measured voltages and currents allows us to derive a total power consumption we can easily confirm with our instrumentation by taking readings at the card's power connectors. In fact, let's start with the measured power consumption values in the following table:

These charts go into more detail on power consumption at idle, during 4K gaming, and under the effects of our stress test. The graphs show how load is distributed between each voltage and supply rail, providing a bird's eye view of load variations and peaks.

Temperature Results

As you can see from the 1070's cooler, Nvidia went with a simple copper heat sink instead of the GeForce GTX 1080's vapor chamber-based solution. Given the price of this card, that's a rather incomprehensible cost-cutting measure.

The beefy frame helps with rigidity, but it was also designed to host thick thermal pads, which dissipate waste heat from the MOSFETs and memory modules.

Unfortunately, Nvidia's cooling compromises have an effect on the temperatures we measured:

Not only does GP104 touch 176°F (80°C), it actually exceeds that value in a closed case. During the gaming loop in an open case, we measured 172°F (78°C) at the GPU package.

We observed similar power consumption during the torture loop, which is due to the fact that this card hits its power target. Not surprisingly, we also see corresponding temperatures.

Sound Results

As a consequence of the higher temperatures, Nvidia's fan has to work harder. During our stress test, the fan speed rose slowly in the beginning due to a lower clock rate. However, after about 30 minutes, it eventually caught up with what we saw during our gaming loop.

The obvious consequence of a spinning fan is noise. Yet, our acoustic measurements are still relatively moderate when the card is idle, despite the sound's slightly snarly character.

Under prolonged load, the noise levels reach almost 42 dB(A). That's still acceptable for a radial fan. At this point, the frequency spectrum becomes rather wide, so the noise doesn't feel too intrusive:

Overall, the radial fan solution is workable. But it's miles away from making this a quiet, gentle card.

MORE: Best Deals

MORE: Hot Bargains @PurchDeals

NEW: Asus ROG Strix GTX 1070

Asus uses its ROG Strix brand to denote the high-end solutions in its portfolio. But where does the ROG Strix GeForce GTX 1070 land in comparison to other companies with premium models?

Technical Specifications

MORE: Best Graphics Cards

MORE: Desktop GPU Performance Hierarchy Table

MORE: All Graphics Content

Exterior & Interfaces

The cooler cover is made of anthracite-colored matte plastic. This does look a bit boring, and it doesn't feel particularly inspiring in-hand, either. Perhaps the card's strengths lie elsewhere.

Weighing in at 1048 grams, Asus' board lands in the heavyweight category. You shouldn't have any trouble securing it into your case, though. A little more critical is its 30cm length. The card's height of 12.5cm is rather average, and its 3.5cm width is typical of all dual-slot designs. A total of three 90mm fans ensure the right amount of airflow and pressure cool the heat sink underneath.

The back of the board is covered by a single-piece plate made of what looks like anodized aluminum. Asus calls this its Aura RGB Lighting Backplate, and it's adorned with a back-lit ROG logo. This backplate makes it necessary to plan for an extra 5mm of clearance behind the card, which may be relevant in multi-GPU configurations.

The card's top sports a Republic of Gamers label with LED background lighting and an eight-pin power connector.

A peek at the end and bottom of the ROG Strix GeForce GTX 1070 reveals that its fins are oriented vertically. They won't allow any waste heat to exhaust out the back. Instead, hot air is pushed from the top and bottom, warming up other components in your case, along with your motherboard underneath. As such, this design decision is rather counterproductive.

The slot plate features five display outputs, of which a maximum of four can be used simultaneously in a multi-monitor setup. In addition to one dual-link DVI-D connector, the bracket also hosts two HDMI 2.0b ports and two DisplayPort 1.4-capable interfaces. No doubt, this is a nod to Rift and Vive owners who need at least one HDMI output for their HMD. The rest of the slot plate is dotted with openings for airflow, though they're not functional due to Asus' fin design.

Board & Components

Asus uses its own circuit board for the ROG Strix GeForce GTX 1070. As far as we can tell, the company doesn't make any concessions in its layout or component choice. However, we also don't see any exclusive traits that'd set this configuration apart.

This card uses eight Samsung K4G80325FB-HC25 modules with a capacity of 8Gb (32x 256Mb). Each chip operates at voltages between 1.305 and 1.597V, depending on the selected clock frequency.

But let's get back to the PCB and power supply. Asus employs an almost oversized 6+1-phase design, wherein the six GPU phases are supplied by uPI Group's uP9511 eight-phase buck controller. The GPU's voltage regulation is implemented using one IR3555 PowIRstage per phase. This chip includes a gate driver, control and synchronous MOSFETs, and a Schottky diode, saving a lot of board space.

The memory's one phase is provided by a uP1666 2/1-phase synchronous buck controller. This phase takes quite a bit of load, so to help it out, Asus' card utilizes two UBIQ Semiconductor QM3054M6 N-channel MOSFETs in parallel on the high side and two QM3056M6 N-channel MOSFETs on the low side.

Asus relies on its homegrown "Super Alloy Power" chokes, which the company says guarantee a particularly stable power supply. What Asus refers to as its SAP technology, MSI calls Military Class. In the end, both are meant to ensure increased durability and longer life expectancy for capacitors. Of course, there is no way for us to verify that claim in a review, but it certainly sounds good.

Asus' ROG Strix GeForce GTX 1070 uses an ITE 8915FN to monitor and control the current. Two familiar capacitors are installed right below the GPU to absorb and equalize voltage peaks.

Power Results

Before we look at power consumption, we should talk about the correlation between GPU Boost frequency and core voltage, which are so similar that we decided to put their graphs one on top of the other. This also shows that both curves drop as the GPU's temperature rises. Asus imposes a relatively low power target of approximately 170W, which in turn causes a relatively frenetic GPU Boost frequency that's primarily limited by the power consumption cap.

After a warm-up run through our variable gaming load, the card's GPU Boost clock rate settles at an average 1946 MHz, down from a starting point of 2015 MHz. Under a more constant load, it falls to an average of 1734 MHz.

The voltage measurements look similar. Readings around 1.062V drop to 1.025V as the board's frequency slides.

Combining the measured voltages and currents allows us to derive a total power consumption we can easily confirm with our instrumentation by taking readings at the card's power connectors.

As a result of restrictions imposed by Nvidia, whereby the lowest attainable frequencies are sacrificed to hit higher GPU Boost clock rates, the power consumption of many factory-overclocked cards is disproportionately high when they're idle. This one can only go as low as 291 MHz. The following table shows what impact that has on our measurements:

These charts go into more detail on power consumption at idle, during 4K gaming, and under the effects of our stress test. The graphs show how load is distributed between each voltage and supply rail, providing a bird's eye view of load variations and peaks.

The 5.1A we measure provides a comfortable margin below the PCI-SIG's 5.5A maximum for a PCIe slot, especially if you're using this card on an older motherboard. Asus only feeds the memory and one GPU phase through the PCIe slot; the other five phases are powered through the auxiliary eight-pin connector.

Temperature Results

Asus' backplate doesn't play an active role in cooling the ROG Strix GeForce GTX 1070. It's mostly aesthetic, though the plate does contribute to the card's structural rigidity.

The ROG logo's back-lighting is achieved using a simple scattering film with an embedded LED. This diode is then plugged into a socket on the card's PCB.

A basic frame on the front cools most of the memory modules, though its shape could be better. While three of the ICs receive some extra cooling thanks to a thermal pad connecting them to the cooler's heat sink, there is one module that gets almost no cooling at all. 

The sink employs a classic heat pipe direct touch design using flattened and sanded pipes. There are four 8mm heat pipes and one 6mm pipe. To dissipate up to 170W, this configuration is more than sufficient.

Reaching temperatures of up to 144°F (62°C), Asus' ROG Strix GeForce GTX 1070 is still in the green. It's running cool enough to ensure the highest possible GPU Boost frequencies allowed by the low power target.

One look at an infrared image of the board reveals the impact of a well-designed cooler on voltage regulation circuitry. As a result of Asus' work, the PCB endures very little thermal stress.

The same holds true when we run our stress test and observe the temperature rising only marginally. This cooler is absolutely sufficient for what it's being asked to do.

Sound Results

Hysteresis is perfectly implemented, allowing the fan curve to leave a positive impression throughout its range. Even after an hour, the three fans spin at less than 1600 RPM.

Registering 37 dB(A) under full load, Asus' card lands in the middle of our pack for noise. But a closer analysis of the frequency spectrum sheds more light on where that reading's peaks appear. Although the three fans do generate some bearing noise, the sound is balanced well. At a purely subjective level, this is less annoying to listen to than the deeper tones produced by other cards with nominally lower dB(A) values.

There's also some electrical noise caused by the voltage converters. But it's in the high-frequency range that most enthusiasts won't be able to perceive.

MORE: Nvidia GeForce GTX 1080 Roundup

MORE: Nvidia GeForce GTX 1060 Roundup

MORE: All Graphics Content

NEW: Gigabyte GTX 1070 G1 Gaming 8G

If there's one thing we've come to expect from Gigabyte's G1 Gaming cards, it's the highest possible performance at a reasonable price. The company's GeForce GTX 1070 G1 Gaming 8G follows that familiar mission with three fans and a fairly bold-looking fan shroud.

Technical Specifications

Exterior & Interfaces

The fan shroud is made of anthracite-colored matte plastic with decorative orange highlights. Weighing in at a modest 797 grams, the G1 Gaming 8G is a true flyweight. A length of 28.2cm should fit easily in most cases, and a height of 11.5cm is about average. The 3.5cm width is typical of dual-slot designs.

A total of three 80mm fans are meant to ensure the right amount of airflow and pressure. The fact that you can't tell these fans are fairly small is due to the card's proportions, which make it look bigger than it actually is.

The back of the board is covered by a single-piece plate made of what looks like anodized aluminum and adorned with a white Gigabyte logo (there is no back-lighting). This backplate makes it necessary to plan for an extra 5mm of clearance behind the card, which may be relevant in multi-GPU configurations.

Up top, the card is branded with a Gigabyte logo back-lit by an RGB LED. There's also a fan-stop indicator and an eight-pin auxiliary power connector.

A peek at the end and bottom of the GeForce GTX 1070 G1 Gaming 8G reveals that its fins are oriented vertically. They won't allow any waste heat to exhaust out the back. Instead, hot air is pushed from the top and bottom, warming up other components in your case, along with your motherboard underneath. As such, this design decision is rather counterproductive.

The slot plate features five display outputs, of which a maximum of four can be used simultaneously in a multi-monitor setup. In addition to one dual-link DVI-D connector, the bracket also hosts one HDMI 2.0b port and three DisplayPort 1.4-capable interfaces. The rest of the slot plate has openings cut into it, which look like they're meant to help with airflow. In this configuration, however, they're not functional due to Gigabyte's fin design.

Board & Components

Gigabyte's PCB is proprietary with some conservative (yet interesting) component choices.

This card uses eight Samsung K4G80325FB-HC25 modules with a capacity of 8Gb (32x 256Mb). Each chip operates at voltages between 1.305 and 1.597V, depending on the selected clock frequency.

One unique addition is the Holtek HT32F52241 32-bit ARM Cortex-M0+ MCU, which Gigabyte uses for controlling the card's RGB effects.

Gigabyte employs an almost oversized 6+2-phase design, wherein the six GPU phases are supplied by uPI Group's uP9511 eight-phase buck controller.

The GPU's voltage regulation is implemented using one Alpha & Omega AON6508 N-channel AlphaMOS on the low side of each phase and one AON6414A on the high side. What's interesting about these two MOSFETs is the low internal resistance across their performance range, as well as the low control values for the gates. It is thus safe to assume that this choice is an attempt to minimize voltage converter losses.

Gigabyte uses Foxconn's Magic coils, which are fully encapsulated and easy to install during automated assembly.

The memory's two phases are supplied by an unmarked PWM controller, which should roughly correspond to a uP1666 2/1-phase synchronous buck controller. It already comes with an integrated bootstrap Schottky diode and gate driver. The high- and low-side MOSFETs are similar to what we found on the GPU's power phases. Just the coils are slightly smaller.

Current monitoring is handled by a triple-channel Texas Instruments INA3221. Two familiar capacitors are installed right below the GPU to absorb and equalize voltage peaks.

Power Results

Before we look at power consumption, we should talk about the correlation between GPU Boost frequency and core voltage, which are so similar that we decided to put their graphs one on top of the other. This also shows that both curves drop as the GPU's temperature rises. Gigabyte uses a power target of ~185W, which in turn causes a relatively frenetic GPU Boost frequency. Short-term drops in clock rate hint that we, at least temporarily, hit the card's power limit.

After a warm-up run through our variable gaming load, the card's GPU Boost clock rate settles at an average 1936 MHz, down from a starting point of 1974 MHz. Under a more constant load, it falls to an average of 1759 MHz.

The voltage measurements look similar. Readings around 1.062V drop to 1.043V as the board's frequency slides.

Combining the measured voltages and currents allows us to derive a total power consumption we can easily confirm with our instrumentation by taking readings at the card's power connectors.

As a result of restrictions imposed by Nvidia, whereby the lowest attainable frequencies are sacrificed to hit higher GPU Boost clock rates, the power consumption of many factory-overclocked cards is disproportionately high when they're idle. This one can only go as low as 240 MHz. The following table shows what impact that has on our measurements:

These charts go into more detail on power consumption at idle, during 4K gaming, and under the effects of our stress test. The graphs show how load is distributed between each voltage and supply rail, providing a bird's eye view of load variations and peaks.

The 3.6A we measure provides lots of margin below the PCI-SIG's 5.5A maximum for a PCIe slot, especially if you're using this card on an older motherboard. Gigabyte only feeds the memory through the PCIe slot; the other six phases are powered through the auxiliary eight-pin connector.

Temperature Results

Gigabyte's backplate doesn't play an active role in cooling the GeForce GTX 1070 G1 Gaming 8G. It's mostly aesthetic, though the plate does contribute to the card's structural rigidity.

The built-in, two-stage heat sink for VRM and coils deserves a positive mention. We also like the fact that the memory modules are cooled directly by this simple design's stabilizing frame.

A direct consequence of this simplicity, unfortunately, is an uninspired and cheap-looking thermal solution. It consists of a finned sink with a thick aluminum base plate that also serves as a mount for two 8mm heat pipes in a flattened direct touch configuration. These pipes don't even cover the GPU entirely. They're hardly sufficient for moving excess heat. It would have been more appropriate for Gigabyte to use three heat pipes instead.

Of course, a maximum GPU temperature of 147°F (64°C) keeps this card in the green. But a safe thermal ceiling isn't attributable to the cooler, but rather Gigabyte's very low temperature target. After hitting that ceiling, the fans are audibly hard at work.

Looking at our infrared images, we see that the GPU package (and thus the PCB) is slightly warmer than the cooled GPU itself. A reading of 172°F (78°C) on the voltage converters is also bearable. But that's not surprising given dedicated cooling.

The same holds true when we run our stress test and observe the temperature rising only marginally. This cooler is merely acceptable, since it's already operating at its limit.

Sound Results

Hysteresis is perfectly implemented; there is no annoying on/off/on loop to worry about. However, permanent fan speeds above 2000 RPM are certainly nothing to celebrate, since they're clearly audible.

Although we've heard worse than 38.3 dB(A) under full load, Gigabyte's fans are still significantly louder than many other GeForce GTX 1070s, and unnecessarily so.

MORE: Nvidia GeForce GTX 1080 Roundup

MORE: Nvidia GeForce GTX 1060 Roundup

MORE: All Graphics Content

Gigabyte GTX 1070 Mini ITX OC

Even the longest can fall short when it comes to mini-ITX enclosures. Suddenly there are other attributes that matter more than just performance, like length, power, and where all of that hot air goes after it's blown off the GPU.

With its GeForce GTX 1070 Mini ITX OC, Gigabyte adapts well to the limitations of small cases, creating a card that's short and still wields Nvidia's powerful GP104 processor. That leaves us with just two questions: how well can the chip be cooled, and how well does Gigabyte manage the trade-off between less surface area and noise output?

This board's price makes it an interesting offering, since Gigabyte deliberately avoids gimmicks and concentrates on the qualities that matter.

Technical Specifications

MORE: Best Graphics Cards

MORE: Desktop GPU Performance Hierarchy Table

MORE: All Graphics Content

Exterior & Interfaces

The cooler cover is made of black plastic with orange lacquered highlights. Weighing just 605g, this card is a featherweight among the other 1070s. A length of just 6 3/4 in (17.2 cm) also makes the GeForce GTX 1070 Mini ITX OC conveniently short. The height (12.5cm from the upper edge of the motherboard) and depth (3.5cm, the same as most dual-slot cards) are perfectly acceptable.

Gigabyte deliberately doesn't use a continuous single-piece backplate; in ITX projects, thicker structures attached to the back of a card often cause collisions with CPU coolers or memory. The smaller plate at the end of the card shouldn't pose any problems, and it does serve a real purpose.

The top of the card bears an unlit Gigabyte logo, and the eight-pin auxiliary power connector sits at the top's back edge. We can appreciate the simple, yet functional design. Gigabyte even held off on eye-catching lighting, since few mini-ITX enclosures are windowed.

Horizontally-oriented fins direct heated air to the back and front of the card. Accordingly, the slot plate has some openings for ventilation.

You get a total of four display outputs, all of which can be used simultaneously in a multi-monitor setup. In addition to the two dual-link DVI-D connectors, there is also one full-sized HDMI 2.0b port and a DisplayPort 1.4-capable interface.

Board & Components

Gigabyte designed an extra-short PCB for this card. The true art is finding a configuration in which the VRMs are arranged in such a way that potential hot-spots are elegantly avoided, despite the lack of space. In practice, Gigabyte's implementation works well.

Like Nvidia's GeForce GTX 1070 Founders Edition, this card uses eight Samsung K4G80325FB-HC25 modules, each able to store up to 8Gb (32x 256Mb).

Such a short card presents issue with placing the memory. On one hand, distance to the GPU cannot be reduced. On the other, the modules must be kept as far away as possible from the voltage regulation circuitry, else they overheat.

The 4+1-phase system relies on the same uP9511P for PWM control as Nvidia's reference card. Gigabyte places this chip on the back of its PCB, though. The memory's one power phase is controlled by a uP1728 on the top side of the board. The converter rails are doubled, a trick that allows each of them to be advertised as two phases, rather than one. Of course, Gigabyte uses an obligatory INA3221 for monitoring current, too.

Gigabyte implemented a voltage regulation solution that allows it to keep the card short without risking thermal issues. Foxconn coils and 6414 high-side MOSFETs are placed on the front of the PCB, while the five gate drivers responsible for controlling the individual phases are banished to the back.

This is also where the 6508 low-side MOSFETs are found. Two MOSFETs are used for each converter circuit, halving the already low internal resistance of the circuits a second time.

The advantage of this setup is that the low-side MOSFETs can be cooled from above by the cooler, while the low-side chips benefit from increased surface area due to the doubling of MOSFETs. This allows them to be cooled by a small screwed-on plate, preventing problematic hot-spots.

Furthermore, two capacitors are installed right below the GPU to absorb and equalize peaks in voltage.

Power Results

Before we look at power consumption, we should talk about the correlation between GPU Boost frequency and core voltage, which are so similar that we decided to put their graphs one on top of the other. This also shows that both curves drop as the GPU's temperature rises.

After warming-up in our gaming workload, the GPU Boost frequency fluctuates between 1778 and 1860 MHz. Under a more constant and taxing load, the clock rates drop significantly to a range between 1584 and 1607 MHz.

The voltage measurements respond similarly. While we observe ~0.975V in the beginning, this value later drops as low as 0.893V. It is easy to tell that Gigabyte had to put clock rates on a diet in order to manage power consumption and thermal output.

Combining the measured voltages and currents allows us to derive a total power consumption we can easily confirm with our instrumentation by taking readings at the card's power connectors.

As a result of restrictions imposed by Nvidia, whereby the lowest attainable frequencies are sacrificed to hit higher GPU Boost clock rates, the power consumption of many factory-overclocked cards is disproportionately high when they're idle. Gigabyte achieves a good compromise, though. Its lowest frequency is 164 MHz, only slightly above Nvidia's reference.

These charts go into more detail on power consumption at idle, during 4K gaming, and under the effects of our stress test. The graphs show how load is distributed between each voltage and supply rail, providing a bird's eye view of load variations and peaks.

Temperature Results

Naturally, power consumption directly affects temperatures. The question of how well Gigabyte's compact GeForce GTX 1070 Mini ITX OC copes with thermal energy can only be answered by looking closely at its cooling solution. The company employs two 1/3-inch (8mm) and one 1/4-inch (6mm) heat pipes made of copper-composite material. They make direct contact with the GPU, accelerating dissipation out through the heat sink's fins.

The small plate for cooling the memory modules is seated on an aluminum heat sink. Three heat pipes are embedded in it, and they're directly connected to the finned cooler.

Gigabyte actively cools the six high-side MOSFETs using the cooler's integrated heat sink. It also draws heat from the coils with thermal pads, and applies slight pressure through them to combat vibration, which we come to know as coil noise. This solution kills two birds with one stone; it eliminates hot-spots and noise simultaneously.

The smaller cooling plate on the back covers all 12 low-side MOSFETs and the gate drivers, dissipating excess heat through the back of the card. As you might imagine, this helps with the small board's thermal challenges quite a bit. Registering 151°F (66°C) on an open test bench, and 154°F (68°C) to 156°F (69°C) in a closed case, the GPU's temperatures are completely acceptable.

A relatively high idle temperature between 118°F (48°C) and 122°F (50°C) is attributable to the card's passive mode. Gigabyte's small cooler doesn't really seem suitable for that purpose. In a compact enclosure, we'd suggest keeping the fan spinning at all times, even slowly. Fortunately, this behavior can be adjusted manually through software.

The infrared images reveal no troublesome hot-spots. Nvidia's GPU, Samsung's memory, and the voltage regulators are cooled well, particularly given a dense PCB.

Temperatures rise a bit during our stress test, but they're not high enough to cause concern. The bottom line is that Gigabyte's concept works.

Sound Results

Since the temperatures during our gaming and torture workloads are similar, the observed fan speeds are about the same as well.

There is no start-up peak in the graph, which would normally indicate the jump from passive to active mode. The installed 90mm fan tops out at 3300 RPM and is stable at speeds as low as 400 RPM. As a result, no tricks are necessary to get it up and running. Gigabyte keeps a good grip on hysteresis as well, and that might be helped by the fact that the fan is activated early, just below 132°F (56°C).

Although the 41.8 dB(A) we measure after a prolonged workload isn't particularly quiet, it's also also not loud enough to bother us. Even Nvidia's GeForce GTX Founders Edition is significantly louder. The high-frequency parts of the VRM still generate quantifiable noise, but they're barely audible. Low-frequency buzzing is rather moderate, and the part that actually is measurable is due to airflow.

Gigabyte does a good job with this small card's cooling; it'd be hard for us to suggest significant improvements.

MORE: Best Deals

MORE: Hot Bargains @PurchDeals

MSI GTX 1070 Gaming X 8G

MSI managed to build a reputation for itself with the Gaming series, and its GeForce GTX 1070 Gaming X 8G strives to live up to high expectations set by those predecessors. The coming employs a conservative default clock rate and a good-enough, low-noise cooler.

There's a higher-end family called Lightning in MSI's portfolio, so it remains to be seen how much the manufacturer held back to create the more mid-range Gaming X version. At the same time, this also isn't MSI's entry-level GeForce GTX 1070. That distinction goes to the simpler Armor 8G and Aero 8G.

We're testing a press sample of the Gaming X with a non-standard firmware, as it turns out. In this version, the OC mode is active by default, resulting in a ~20 MHz-higher core and GPU Boost clock rate. In practice, the differences between this card and the retail one are negligible, so we didn't re-flash the BIOS for our tests. Instead, we switched back to the standard mode via vendor-supplied software. Thus, the GPU's frequency in our benchmarks is still equivalent to the retail model.

Technical Specifications

MORE: Best Graphics Cards

MORE: Desktop GPU Performance Hierarchy Table

MORE: All Graphics Content

Exterior & Interfaces

MSI's fan shroud is made of relatively thin plastic, though it's mounted on a large frame for improved stability. Weighing in at 1080g, this card isn't particularly light. It's 28.4cm long, 13.5cm tall, and 3.5cm wide, similar to most dual-slot cards. The rotor blades of the double ball bearing fans have a diameter of 95mm.

The back of the board is covered by a single-piece plate with some holes for ventilation, but it doesn't contact the PCB through thermal pads and consequently does nothing for cooling. Unfortunately, this backplate makes it necessary to plan for an extra 5mm of clearance behind the card, which may be relevant in multi-GPU configurations.

Using the card without its backplate proved difficult due to integrated spacers. Furthermore, the screws attaching the backplate have threaded portions that are too short to secure the frame and VRM cooler once the plate is gone. That's why we had to abstain from running the card without its backplate, limiting our view during the IR temperature testing.

The top of the card is dominated by an illuminated MSI logo and three visible heat pipes (two 6mm and one 8mm). Six- and eight-pin auxiliary power connectors are positioned at the end of the card and rotated by 180°. MSI's black and red color scheme should be familiar by now, similar to the card's shape.

An open end reveals horizontally-oriented fins, which channel airflow towards the card's front and back. This is an advantage since a lot of heated air exhausts out the slot cover. What remains isn't directed out the top, close to your CPU. Although we like centrifugal coolers able to blow everything out the back, this is a next-best solution.

The slot plate features five display outputs, of which a maximum of four can be used simultaneously in a multi-monitor setup. In addition to one dual-link DVI-D connector, the bracket also hosts one HDMI 2.0b port and three DisplayPort 1.4-capable interfaces. The rest of the slot plate is dotted with openings for airflow.

Shedding that often-unused DVI connector would have made room for even more ventilation, which may have helped the cooling solution's effectiveness. Ah well.

Board & Components

A glance at the board reveals a well-arranged layout. It uses eight Samsung K4G80325FB-HC25 modules, each able to store up to 8Gb (32x 256Mb). Each chip operates at voltages between 1.305 and 1.597V, depending on the selected clock frequency.

The 8+2-phase power subsystem relies on the sparsely documented µP9511P for PWM control. However, unlike Nvidia's Founder Edition card, the controller is placed on the front rather than the back. Since this controller can't communicate directly with the VRM's phases, MSI utilizes gate drivers to talk to the dual-channel MOSFETs. And because MSI is using dual- rather than single-channel MOSFETs, it can use space on the PCB more efficiently. The company also goes with Super Ferrite Chokes, which are a bit classier than what you typically find on reference designs.

Two capacitors are installed right below the GeForce GTX 1070 Gaming X's GPU to absorb and equalize voltage peaks.

Power Results

Before we look at power consumption, we should talk about the correlation between GPU Boost frequency and core voltage, which are so similar that we decided to put their graphs one on top of the other. This also shows that both curves drop as the GPU's temperature rises.

After warm-up, GPU Boost frequencies fall as low as 1962 MHz under load. This is mirrored by our voltage measurements. While we observed up to 1.062V in the beginning, just like Nvidia's Founders Edition, that value drops to 1.05V later on.

Combining the measured voltages and currents allows us to derive a total power consumption we can easily confirm with our instrumentation by taking readings at the card's power connectors.

As a result of restrictions imposed by Nvidia, whereby the lowest attainable frequencies are sacrificed to hit higher GPU Boost clock rates, the power consumption of many factory-overclocked cards is disproportionately high when they're idle. This one can only go as low as 253 MHz. The following table shows what impact that has on our measurements:

These charts go into more detail on power consumption at idle, during 4K gaming, and under the effects of our stress test. The graphs show how load is distributed between each voltage and supply rail, providing a bird's eye view of load variations and peaks.

Temperature Results

As it pertains to cooling, MSI has a lot to offer. Its sturdy frame covers almost all of the PCB not already covered by heat sinks for the VRMs.

Thermal pads between the memory modules and frame, along with the voltage regulators and cooling plate, provide the interfaces needed to transfer heat away from those areas. However, we would have liked thinner and more efficient design, since fan speeds (and thus airflow) are rather low.

The massive cooler has horizontally-oriented fins as well as four nickel-plated 1/4-inch (6mm) copper heat pipes, plus an 8mm one. The direction they face is more or less irrelevant since they consist of sintered composite material, relagating unreliable axial grooves or meshes to ancient history. MSI furthermore relies on a massive nickel-plated heat sink.

A rather restrictive 149°F (65°C) temperature target does result in the fans reacting faster with a shorter start-up time. But this is unfortunately foiled by their throughput and maximum required speed.

Since we couldn't test the GeForce GTX 1070 Gaming X 8G without its backplate, we had to take our measurements with that cover in place. On the bright side, one of the plate's ventilation openings is right above a memory module positioned between the VRMs and GPU. The measured 176°F (80°C) is acceptable.

Temperatures actually dropped during our stress test, since the card throttled back its clock rate, lowering power consumption.

Sound Results

Since the temperatures during our gaming and torture workloads are similar, the observed fan speeds are about the same as well. The start-up behavior and well-implemented hysteresis, which prevents multiple on/off cycles once the fans start spinning, is clearly visible.

In addition, the start-up speed is chosen in such way that the fans will certainly continue to start reliably, even as they age. The same goes for when the fans turn off.

When the card is idle, there is no noise thanks to its semi-passive mode. Naturally, there's no need to take acoustic measurements.

Running at full load barely raises the noise level to 33 dB(A). This value doesn't go any higher, even during our torture test. The card is so quiet that we were able to measure >5 kHz frequencies coming from the voltage regulators, which most people would barely be able to hear. It's only when the GeForce GTX 1070 Gaming X 8G cranks out frame rates in the hundreds that you hear the VRMs chirp a little louder.

It seems like the cooler is almost overkill for its job. On the other hand, we prefer being safe to sorry. MSI obviously does a good job.

MORE: Best Deals

MORE: Hot Bargains @PurchDeals

Palit GTX 1070 GameRock Premium Edition

JetStream, Super JetStream, GameRock, and GameRock Premium Edition: Palit sure does provide a lot of GeForce GTX 1070-based options. The card we're testing is just as bulky as its name thanks to the oversized cooler.

Despite all of the cooling headroom you'd seemingly get, this version does have issues with hysteresis, causing the fans to start and stop during warm-up. Unfortunately, the situation isn't any better, even after downloading an available firmware update. Consequently, we're putting further updates to this piece on hold until we're offered a real solution.

Just like the GTX 1080-based version, this card looks massive at first glance. But we'll have to test how much of that bold appearance translates into real-world benefits.

Technical Specifications

MORE: Best Graphics Cards

MORE: Desktop GPU Performance Hierarchy Table

MORE: All Graphics Content

Exterior & Interfaces

The fan shroud is made of relatively thick, white plastic. The top and front are decorated with metallic blue and  brushed metal highlights.

A weight of 1053g makes this card about 180g lighter than the 1080-based version. However, with its length of 28.7cm, height of five inches (12.8cm), and two inches (5.2cm) of width, it boasts the same dimensions as its higher-end counterpart and also spans three slots. The two massive 100mm fans with rotor diameters of 96mm further emphasize the card's size.

Flip the card over and it's covered by a one-piece backplate without any openings for ventilation. It is decorated with a highly visible GameRock logo and requires an additional 5mm for clearance. Since there are no thermal pads between the plate and PCB, the backplate serves decorative purposes only. While it is perfectly possible to use the card without this plate, removing it requires taking off other components, likely voiding Palit's warranty.

The top of the card is dominated by a centered, brightly lit Palit logo. An eight-pin auxiliary power connector is positioned at the board's end and rotated 180°. This isn't a humble product by any means; it stands out and wears its heft proudly.

At its end, the card is completely closed off, which makes sense since the fins are positioned vertically and won't allow any airflow in that direction anyway.

The slot bracket features five display outputs, of which four can be used simultaneously in multi-monitor setups. In addition to one dual-link DVI-D connector, which doesn't loop through any analog signal, you also get one HDMI 2.0b port and three DisplayPort 1.4-capable interfaces. The rest of the slot plate is peppered with openings that look like they're meant to facilitate airflow. However, they're more decorative than functional.

Board & Components

The board looks clean enough, similar to the 1080-based version. It uses eight Samsung K4G80325FB-HC25 modules, each able to store up to 8Gb (32x 256Mb). Each chip operates at voltages between 1.305 and 1.597V, depending on the selected clock frequency. However, caution should be advised! Some manufacturers have switched to Micron memory modules, which is recognizable in the BIOS naming scheme (86.04.26.xx.xx versus Samsung's 86.04.1E.xx.xx).

The 8+1-phase system, like Nvidia's reference cards, relies on the sparsely documented uP9511P for PWM control. Also like Nvidia's own implementation, the controller finds a home on the back of the PCB. All eight of the GPU's phases are realized using this component, which is actually designed as a 6+2-phase jack-of-all-trades.

The DC/DC voltage converters' dual-channel MOSFETs are controlled directly, as these eight SiC632s are so-called driver MOSes. They combine the actual power MOSFETs for high-side and low-side, as well as the gate driver and Schottky diode on one chip. This is certainly more cost-efficient and a plus for compact designs, especially when a large number of voltage converter circuits are in play.

In contrast to Palit's GTX 1080 model, this board's memory gets power from one phase (instead of two). It's controlled by the same undocumented chip used on Nvidia's reference board, which should be almost identical to the well-known 1728. A dual N-Channel model is used for the MOSFETs, which combines both high-side and low-side.

The Foxconn coils are middle-class. Depending on the layout, they operate more or less quietly. And as is often the case, a well-known INA3221 handles current monitoring.

Two capacitors are installed right below the GPU to absorb and equalize peaks in voltage.

Power Results

Before we look at power consumption, we should talk about the correlation between GPU Boost frequency and core voltage, which are so similar that we decided to put their graphs one on top of the other. This also shows that both curves drop as the GPU's temperature rises.

The graphs clearly show that the GPU Boost frequency after warm-up and under load falls from an excellent 2076 MHz to a still-good 1975 MHz (and sporadically a little lower). It is also apparent that voltage follows the sinking clock rates. While we measured up to 1.062V in the beginning, this value later drops as low as 0.901V.

Combining the measured voltages and currents allows us to derive a total power consumption we can easily confirm with our instrumentation by taking readings at the card's power connectors.

Since manufacturers sacrifice the lowest possible frequencies to gain an extra GPU Boost bin due to Nvidia's restrictions, the GTX 1070 GameRock Premium Edition's power consumption is slightly higher at idle. Palit sets the first GPU Boost step at 316 MHz.

These charts go into more detail on power consumption at idle, during 4K gaming, and under the effects of our stress test. The graphs show how load is distributed between each voltage and supply rail, providing a bird's eye view of load variations and peaks.

Temperature Results

Of course, power dissipated as waste heat needs to be dealt with as efficiently as possible. So, we start by looking at the backplate, which doesn't do any real cooling and instead leaves that job to Palit's 2.5-slot thermal solution. The entire structure matches what you get on the GTX 1080 card. It appears almost decadently oversized given how much less power a 1070 consumes. On the other hand, this isn't a bad problem to have. We'd rather have too much cooling than not enough.

A copper sink moves heat away from the GPU and spreads it through a total of five pipes (three 8mm and two 6mm). Palit chose to orient the sink's fins vertically, which results in short, straight 8mm pipes that work more efficiently. The two smaller pipes don't do much except provide additional area to support the transport of heat away from the sink and towards the cooler's edges.

The performance of this truly monstrous cooler leaves little to be desired. Since the temperatures only get to 151°F (66°C) during our gaming loop (154°F/68°C in a closed case) and 151°F (66°C) as well during our stress test (158°F/70°C in a closed case), the fans only need to run at low power, which should have a positive impact on the measured noise level.

The transfer of heat away from the VRMs works perfectly, despite low fan speeds and very little airflow. That massive cooler and its endless fins works wonders.

During our stress test the temperatures do rise a bit at the hottest spot, despite low average power consumption. All other areas remain at perfectly safe values, though.

Sound Results

Let's talk a bit about the one single quality that dominated our impression of this graphics card: its noise level. While some folks might enjoy the deep roar of a well-oiled machine, we the sound of silence from our graphics cards. When it's disturbed, the culprit is usually a fan or sometimes the VRM's coils. But a maximum of 1000 RPM for the two fans should be no reason to declare a state of acoustic emergency.

To examine this behavior in more detail, we need to take a closer look at the fan curve, which unfortunately reveals an unpleasant surprise. Since the fans generally start late and keep quiet, it takes diligent measurement of tachometer signals and PWM values to detect poor hysteresis. Palit confirmed this for us, even.

When the card is idle, no noise is measurable due to its semi-passive mode. We thus abstained from trying to take any readings.

The values we measured under load are blissful, and the machine is a purring kitten. A result of 35.1 dB(A) is good considering the temperatures involved. Only the mid-range coils tend to stick out a bit. If it wasn't for their audible chirp disturbing the calm of night, you might even question whether the card was running at all.

MORE: Best Deals

MORE: Hot Bargains @PurchDeals

Introduction & Overview

AMD's Radeon RX 480 launched last June (if you missed our in-depth AMD Radeon RX 480 8GB Review, be sure to check it out). In the months that followed, we covered new GPUs and embarked on a sequence of in-depth round-ups across Nvidia's Pascal architecture. We always planned to circle back on Polaris, though.

This proved more difficult than we anticipated, as availability from AMD's board partners started off spotty. Fortunately, supply improved and mainstream enthusiasts now have plenty of options to choose from. Some models sport AMD's reference cooler, while others employ custom thermal solutions. Some models come with 4GB of GDDR5 memory and others wield 8GB. Some models sell for as little as $175; the top end spans beyond $260. Now it's our job to figure out which designs work better than the rest.

At launch, the Radeon RX 480 8GB trailed just behind GeForce GTX 970, and later came to follow after the GeForce GTX 1060. But back then, most of our benchmarks were DirectX 11-based. The introduction of newer titles optimized for DirectX 12 painted Polaris in a more attractive light. Today, the two cards trade blows, with AMD's solution generally selling at a lower price.

MORE: Nvidia GeForce GTX 1060 Roundup

But this round-up definitely isn't a comparison between Radeon RX 480 and GeForce GTX 1060. In this piece, AMD's Radeon gets the spotlight. Our aim is to explore, in detail and with a great many measurements, the differences between individual implementations of AMD's Radeon RX 480.

As always, there's very little each manufacturer can do to affect core and memory clock rates. But the opposite is true when it comes to their interpretations of power limits, on-board components, and cooling. That's where we'll focus most of our time and equipment. In fact, we'll present data with these cards on an open test bench and in a closed chassis. In some cases, the differences are significant!

At least for the time being, we have five cards to discuss. More are on their way, and we have updates coming for some of our other round-ups as well.

MORE: Best Graphics Cards

MORE: Desktop GPU Performance Hierarchy Table

MORE: All Graphics Content

MORE: Best Deals

MORE: Nvidia GeForce GTX 1080 Roundup

MORE: Nvidia GeForce GTX 1070 Roundup

1080p Game Benchmarks

We test every card after a suitable warm-up period to avoid unfair differences in boost frequencies. All benchmarks are run six times; the first one is used to get the GPU hot again.

These cards are all press samples operating at the same settings as retail models in our best effort to ensure one vendor doesn't get a leg up on another using non-representative clock rates.

The following two pages cover the resolutions we consider relevant to AMD's Radeon RX 480: 1920x1080 and 2560x1440. Our benchmark suite includes games known to favor AMD and Nvidia, demonstrating the full range of potential performance differences.

For Honor

Grand Theft Auto V

Mafia 3

Sniper Elite 4

Watch Dogs 2

The Witcher 3

MORE: Best Graphics Cards

MORE: Desktop GPU Performance Hierarchy Table

MORE: All Graphics Content

1440p Game Benchmarks

The following charts cover 2560x1440:

For Honor

Grand Theft Auto V

Mafia 3

Sniper Elite 4

Watch Dogs 2

The Witcher 3

All of the factory-overclocked cards offer similar performance, more or less. Apparently, these cards already operate so far above their sweet spot that performance no longer scales linearly with clock rate, and thus the price for even a slightly higher frame rate is a disproportionate increase in power consumption. Thus, to find the "best" Radeon RX 480 for your purposes, compare power and cooling, rather than the gaming results. Especially for the RX 480, a more modest approach pays off.

MORE: Best Graphics Cards

MORE: Desktop GPU Performance Hierarchy Table

MORE: All Graphics Content

Power, Sound & Temperature Benchmarks

Power Consumption

We begin by comparing the power consumption of each card in our gaming loop, stress test, and at idle. Depending on the power targets specified by each manufacturer, we sometimes measured substantial differences. We're ignoring the decimal places in our bar graphs, since those values are too small and would be within measurement tolerances.

We also found that some cards with lower power targets started to throttle during our stress test, resulting in lower power consumption numbers. This did not, however, have a negative impact on general gaming performance for any of the tested cards, as the stress test merely represents a worst-case scenario.

Sound

In the following comparison, we summarize the maximum results in our gaming loop as a bar graph - check out the frequency spectrum on the individual product pages, too.

Since these cards employ a semi-passive mode, their fans don't spin at idle. Consequently, there was no reason to collect measurements in that state.

Temperature

In two galleries we compare temperatures on the open test bench and inside a closed case. We start with measurements from the GPU:

The VRMs are just as important as the GPU since they play an even more active role in heating the PCB. Again, you'll find more detail on the individual products pages.

MORE: Best Graphics Cards

MORE: Desktop GPU Performance Hierarchy Table

MORE: All Graphics Content

Reference AMD Radeon RX 480

Aside from the fact that AMD is first alphabetically, we also wanted to cover the company's card first for another reason. Reference designs often have to defend against the perception that they're not as good as what follows from board partners. And in this case, that's not entirely true.

AMD didn't do a bad job with this card. That much is evident after testing a number of third-party alternatives. In fact, in a closed case, the centrifugal cooler that exhausts waste heat may even outperform some of the factory-overclocked boards with axial fans in certain metrics.

In an effort to generate results from a newer sample, we got our hands on another reference card and repeated the measurements from our launch coverage. Even if the results differ a bit, we considered it more fair to test with a more recent chip than one of the first, especially since our equipment changed a bit as well.

Specifications

MORE: Best Graphics Cards

MORE: Desktop GPU Performance Hierarchy Table

MORE: All Graphics Content

Exterior & Interfaces

The card weighs just 24.1 ounces (685 grams), is 24.2cm long, 10.5cm tall, and 3.5cm wide. At first glance, its dimensions look similar to AMD's Radeon R9 Fury Nano. However, a closer look reveals several cost-cutting measures compared to the previous generation's highest tier.

The black plastic body is one piece, even if the four supposed Allen screws suggest otherwise. They're merely a ruse to hint that this is a multi-part design.

Around back, we see that the actual PCB is only 18cm (or eight inches) long. We've seen this sort of overhanging radial fan and bottom-facing air intake before from Nvidia.

Up top, next to the Radeon logo, the card features a single six-pin auxiliary power connector. It's barely ample, as we saw in our launch review and AMD Radeon RX 480 Power Measurements Repeated And Clarified. There should be an eight-pin connector in this spot, and we'll go into more detail about that shortly.

A fully closed end encourages air to flow from the fan out through the front bracket, keeping it from heating up any of your other components.

AMD eliminates DVI output from the output bracket to make more room for ventilation holes. You can still attach up to four monitors simultaneously, though, via three DisplayPort 1.4-ready connectors and an HDMI 2.0 interface.

Board & Components

A look at the PCA reveals that the VRMs are moved near the output bracket. We consider this to be counterproductive, and we're not sure why AMD went that route. The Ellesmere GPU, bare of any identifying markings, is situated diagonally in its package, similar to older Radeon graphics cards.

AMD goes with Samsung GDDR5 memory modules (model number K4G80325FB-HC25). Each one has a capacity of 8Gb (32x 256Mb). And depending on clock rate, they can operate at voltages between 1.305V and 1.597V. Their ceiling is 2000 MHz, similar to what we find on Nvidia's GeForce GTX 1070.

Zooming in for a closer look at the board, we find International Rectifier's IR3567B dual-output 6+2-phase PWM controller. AMD started using this buck controller as far back as the Radeon R9 290. It's the heart of the company's power supply and is well-documented. That makes programming control and overclocking utilities for the RX 480 much easier.

Six phases are used to supply the Ellesmere GPU with power. Unlike Nvidia, which uses one dual N-channel MOSFET for the job, AMD relies on two single N-channel MOSFETs per phase. There's one MDU1514 (30V, 66.3A, 6mΩ) and one MDU1511 (30V, 100A, 2.4mΩ), controlled by an IR CHL8510 high-performance gate driver.

The seventh phase supplies power to the memory modules, and in contrast to Fiji-based cards, it uses the motherboard slot's 3.3V rail. This makes the design significantly simpler, especially since the modules don't consume much power. In this one regard, we're glad the RX 480 doesn't follow the Radeon R9 Fury X/Fury/Fury Nano's lead in hosting the gate driver on the PCB's back side, directly behind the MOSFETs, creating serious hot-spots.

Power Results

Even after adjusting the drivers and enabling AMD's compatibility mode, the reference card's power consumption is still above its 150W TDP, though just barely.

During our gaming loop, the card consumes 157W. We measured just above 158W during the stress test. AMD's Radeon RX 480 simply can't go much higher.

We mention short load peaks only as a side note (see the light-grey bar in the chart below), since they're rarely relevant in practice. Nevertheless, they illustrate the importance of deploying a reasonably-sized secondary side for the power supply (low-impedance caps).

The following two graphs represent two minutes captured from the gaming loop and stress test. These are what our average power consumption figures are calculated from:

The next two graphs map current measurements at each of the supply rails, corresponding to the power consumption levels shown in the diagrams above:

What we see is an almost perfect match. Think back to our Radeon RX 480 launch coverage: it was a nonstandard load on the motherboard slot's 12V rail that forced AMD to publish its driver-based fix. Thanks to the compatibility mode function, everything now seems to be in order.

If you own an older motherboard, we still recommend using AMD's compatibility mode to avoid potential problems. Owners of newer platforms have nothing to worry about, though.

Temperature Results

Pulling the shroud off is easy after removing six screws on the side. Under the cover, you'll find a simple heat sink with fins, a larger frame for stabilization and cooling, and a radial fan. We've seen the fan before, though this time its maximum rotational speed is set at 5000 RPM.

From this angle, it's clear that the Radeon RX 480 employs direct heat exhaust cooling, where waste heat is dissipated through the sink and blown out through the slot bracket.

The heat sink is a simple piece of extruded aluminum with an embedded copper core to improve thermal conductivity above the GPU. A strip glued to the top of the fins ensures that the air doesn't leak out above the cooler, but instead passes through it entirely.

A frame underneath the heat sink serves double duty, helping cool the memory modules and MOSFETs. Fins in the frame help increase its surface area somewhat.

The cooling performance that AMD achieves isn't bad. During our gaming loop, the Radeon RX 480 sustains an average frequency of about 1225 MHz, and only sporadically drops by a few megahertz. The reasons for this are tied to an imposed power limit, which makes itself known at certain points under taxing workloads.

As we've come to expect from direct heat exhaust-based graphics cards, temperature measurements in an open test bench and closed case are very similar. Only the GPU is marginally warmer, but this could very well be caused by a higher ambient temperature inside of our enclosure.

The values are again similar during our stress test, showing minimal difference between the open and closed setups. In a case, the slightly higher temperature comes from a minor increase in power consumption.

Regardless of whether you're in an airy tower or a cramped form factor, AMD's reference Radeon RX 480 sports sufficient cooling. The only worrisome measurement we collected came from a memory module that hit 181°F (83°C), which is just below the maximum specified by Samsung.

Sound Results

Often, the trade-off for blowing hot air out of your case is a less pleasant-sounding fan. How does the RX 480's thermal solution sound as it facilitates consistent cooling performance?

Once the RX 480 reaches its maximum temperature in our gaming loop, the 41.9 dB(A) reading we collect is—up to the decimal point—identical to what we saw in our launch coverage. That's no surprise, given the same fan speeds. But it's always good to see consistency, even months later.

At idle, a measurement of 31.1 dB(A) is slightly lower than what we saw previously. At low speeds, however, the radial fan does sound rather unpleasant, almost like growling. The spectrum contains some dominant low-frequency parts.

This reference card isn't bad. We've certainly seen reference-class hardware from both AMD and Nvidia that fared worse. Thanks to a driver update that rectified the issues we identified in our launch coverage, AMD's Radeon RX 480's power supply is much better balanced. It's even possible to optimize this card further through WattMan, since chip quality seems to have improved and the voltage presets are higher than they need to be.

AMD's cooling performance is quite convincing, and the fan is acoustically tolerable, despite high RPM values. In fact, this reference implementation really has no reason to shy away from comparisons with partner boards. Can it outshine any of them, though? That remains to be seen.

MORE: Nvidia GeForce GTX 1080 Roundup

MORE: Nvidia GeForce GTX 1070 Roundup

MORE: Nvidia GeForce GTX 1060 Roundup

Asus ROG Strix RX 480

Thanks to its three fans, Asus' ROG Strix RX 480 looks impressive. Moreover, we know from testing it that this board represents a well-executed compromise between performance and practicality. Unlike most other Radeon RX 480s out there, Asus really does the Ellesmere GPU justice.

A factory overclock to "only" 1310 MHz and a moderate power target make it clear the company isn't going all-out for speed. That'd only earn this card a few extra frames per second anyway. Instead, Asus sticks with its traditional strengths.

Specifications

MORE: Best Graphics Cards

MORE: Desktop GPU Performance Hierarchy Table

MORE: All Graphics Content

Exterior & Interfaces

Weighing in at 36.7 oz (1041 grams), the ROG Strix RX 480 is distinctly heavy. It also qualifies as a dual-slot solution at 30cm long, 12.6cm tall, and 3.5cm wide. Still, though, those dimensions could spell trouble in small enclosures, particularly due to the card's length. Because of the backplate, you must plan for an additional one-fifth of an inch (5mm) of clearance to avoid collisions with your CPU's cooler, again complicating installations in compact cases.

Visually, the card employs a typical Strix design with an LED-lit logo as part of its backplate. Enthusiasts with windowed cases and an affinity for flashiness will appreciate the eye candy, to be sure.

The cooler cover is made of plastic. In our opinion, it looks alright. At least the shroud feels strong enough, given the card's heft. Three 8.5cm fans push plenty of air through the heat sink, even at relatively low speeds. Given the thermal solution's relatively low profile, a lot of static pressure isn't really needed anyway.

Asus' DirectCU III cooler employs vertically oriented fins, which makes sense for a three-fan design. The top of the card features an eight-pin auxiliary power connector. Other than that, it's as unremarkable as the bottom. If you want to see more of the actual cooler, you'll have to remove the shroud. Challenge: accepted.

At its end, the card features two PWM-controlled fan headers. These could be used to connect a pair of case fans that are only activated if high temperatures are detected. While that'd be a good idea, most cases unfortunately only have three-pin voltage-controlled fans installed up front. These won't work. Worse, Asus neglects to mention the maximum load the headers accommodate.

The slot bracket is loaded with outputs. You get one DVI-D port, two DisplayPort 1.4-ready connectors, and a pair of HDMI 2.0 interfaces. This doesn't leave a ton of room on the bracket, but Asus does find space for a number of cut-outs. Theoretically, these would help exhaust waste heat. Because the heat sink's fins run vertically, though, the holes serve no real purpose.

Board & Components

With the backplate, stabilization frame, and cooler removed, we get a better look at the PCA. Asus places the voltage regulation circuitry for the GPU and memory on the right side, leaving the left almost completely empty.

Like AMD, Asus uses eight Samsung GDDR5 memory modules (model number K4G80325FB-HC25). Each one has a capacity of 8Gb (32x 256Mb). And depending on clock rate, they can operate at voltages between 1.305V and 1.597V. Their ceiling is 2000 MHz, similar to what we find on Nvidia's GeForce GTX 1070.

The GPU is fed by six phases provided by an International Rectifier ASP1300. What makes this programmable PWM controller special are its overclocking capabilities. Using the SMBus, an extensive command set can be utilized, including access to a VID of up to 2.3V. Features like a freely configurable maximum voltage, VID override, individual load-line calibration, and OVP/OTP stand out as unique compared to what we normally have access to.

International Rectifier's IR3555 is a good match, containing a buck gate driver IC co-packed with control and synchronous MOSFETs, plus a Schottky diode.

Asus relies on its in-house Super Alloy Power feature to describe the chokes and capacitors, which it says facilitate cooler operation and a longer life expectancy. Of course, we can test for temperature, but don't have a metric for capacitor longevity yet.

In contrast, the memory's single power phase is a more conservative choice. A second, seemingly similar voltage converter provides power for peripherals and controllers; it features a somewhat simpler smoothing.

Here, Asus uses the uP1540 as a PWM controller. For each of the two phases, it controls one QM3054 (30V, 4.8mΩ, 97A) N-channel MOSFET on the high side and two QM3056 (30V, 4.2mΩ, 103A) N-channel MOSFETs on the low side.

Power Results

The almost 20W idle result doesn't impress us, but we can see it's caused by rather generous voltage settings and a high frequency of 300 MHz. Hopefully, AMD improves this behavior in its next generation of graphics cards.

During our gaming loop, the ROG Strix RX 480 consumes approximately 175W. That's only moderately higher then what we saw from the reference design. Meanwhile, our stress test reports 173W, which is imposed by a 180W power limit.

We mention short load peaks only as a side note (see the light-grey bar in the chart below), since they're rarely relevant in practice. Nevertheless, they illustrate the importance of deploying a reasonably-sized secondary side for the power supply (low-impedance caps).

The following two graphs represent two minutes captured from the gaming loop and stress test. These are what our average power consumption figures are calculated from.

The next two graphs map current measurements at each of the supply rails, corresponding to the power consumption levels shown in the diagrams above.

Unlike some of the other cards we're looking at in today's round-up, these results should put your mind at ease. A total of just 3.8A is comfortably low compared to the allowed 5.5A. Only the memory's power comes from the motherboard slot; the GPU's six phases get their power from an eight-pin auxiliary connector.

Temperature Results

Asus' cooler is fairly conservative, consisting of a heat sink, frame, and backplate.

Only the frame helps draw heat away from the memory. Unfortunately, its coverage of the modules is partial. Whether or not that's ample remains to be seen. At least the memory also receives some airflow from the fans.

Lacking thermal pads, the backplate doesn't help with cooling. As mentioned, though, the plate's back side does sport an LED for illuminating the logo.

The cooler itself consists of two parts. The larger piece is positioned atop a solid block of aluminum that holds one 6mm and four 8mm nickel-plated heat pipes. The latter are flattened over the GPU for better contact.

We do like that the VRM heat sink is built into the cooler, leveraging its large array of fins. Asus could have done better, though, with a multi-level solution touching the chokes as well.

Either way, the thermal performance we observe is superb. As the auto enthusiast saying goes, there's no replacement for displacement, or in this case heat sink surface area. With a maximum of 149°F (65°C) in a closed case and 147°F (64°C) on an open test bench, the GPU stays nice and cool, even during our stress test. The clock rate does drop a bit, but that's related to Asus' power limit and not the processor's temperature.

Still, an average of 1280 MHz is pretty good. At this same point, competing cards consume up to 50W more for an extra 50 MHz. That small of a gain in the face of much higher power use just doesn't make sense.

Remember when we mentioned that Asus' frame only partially covers the memory? On our test bench, those modules heat up to about 176°F (80°C), while in a closed case the temperature stays just below 181°F (83°C). Although this is still within Samsung's specs, a larger frame would have helped ensure better results.

Topping out at 183°F (84°C), the voltage regulation circuitry has headroom to spare.

A maximum value of 183°F (84°C) in a closed case shows the hottest module right at the limit of what's acceptable. This probably isn't a dire situation, though. After all, we're deliberately invoking a worst-case scenario. A real-world workload won't look as bad.

Asus' DirectCU III cooler shows that it's generally possible to ensure roughly the same temperatures, regardless of whether you're measuring on an open test bench or in a closed case. This becomes an advantage for Asus as we test competing cards that start gasping for cool air in an enclosed environment.

Sound Results

But how does the company go about achieving nearly identical temperatures in our gaming loop and stress test, regardless of whether its card is on an open chassis or in a closed case? A look at the fan speeds reveals they (unsurprisingly) went from 500 to 600 RPM when we buttoned everything up into an enclosure. You might consider that to be problematic. However, it looks like this behavior is by design.

An outstanding measurement of 37.8 dB(A) under a gaming load sets the bar high for competing cards, especially since our readings are even lower on an open test bench. The noise generated is almost free of the low-frequency peaks you'd hear from poorly-manufactured fan bearings. Coil whine is almost non-existent, too. Instead, the noises integrate well together, yielding an acoustic work of art.

By deliberately not trying to get every last bit of frequency out of Ellesmere, Asus came up with a well-cooled card that also operates quietly. It doesn't matter if you use the ROG Strix RX 480 on an open test bench or inside of a closed case, the readings we recorded are just right.

In comparison to the most aggressively tuned competition, Asus loses out by a couple of frames per second at worst, despite its conservative power consumption. But the performance loss isn't noticeable in practice.

What you will notice is the subtle fan work that results in a pleasant-sounding card. Perhaps the worst thing we can hold against the ROG Strix RX 480 is its length. Smaller cases may have a difficult time accommodating such a long fan shroud. Of course, you also need to be alright paying ~$240 for an 8GB Radeon RX 480. With that said, if you're dead set on Polaris, Asus' card is one of the top options out there.

In our opinion, this is the only RX 480 worthy of an Editor's Choice award thanks to its successful balance of efficiency, gaming performance, and cooling. There are faster cards in our test field, but none of them combine the other qualities we consider to be important.

MORE: Nvidia GeForce GTX 1080 Roundup

MORE: Nvidia GeForce GTX 1070 Roundup

MORE: Nvidia GeForce GTX 1060 Roundup

HIS RX 480 IceQ X² Roaring Turbo

With its Roaring Turbo implementation, HIS pushes GPU clock rates up to 1338 MHz, giving the two fans a challenging task in the process. So, does Roaring refer to this card's performance, noise, or both? We're interested to see if that frequency can be maintained under our taxing workloads.

HIS is now a subsidiary of Pine, just like XFX. So, during the course of our analysis, we'll have plenty of opportunities to draw parallels to the XFX equivalent of this card. That's good news, since the RX 480 IceQ X² Roaring Turbo isn't readily available to our U.S.-based readers. The two boards aren't exact clones of each other, though. HIS does take some liberties that we'll explore in depth.

Specifications

MORE: Best Graphics Cards

MORE: Desktop GPU Performance Hierarchy Table

MORE: All Graphics Content

Exterior & Interfaces

Weighing 32 ounces (906 grams), the RX 480 IceQ X² Roaring Turbo is no lightweight. It still fits within a dual-slot form factor though, measuring 28.9cm long, 13.5cm tall, and 3.5cm wide. Because of the backplate, you must plan for an additional one-fifth of an inch (5mm) of clearance to avoid collisions with your CPU's cooler, complicating installations in compact cases.

The fan shroud is made of light metal and does exude higher quality than some of the plastic-covered competition, even if it makes the card unnecessarily tall. It could have been 2cm shorter, improving its compatibility with a greater number of cases. A gold-colored decal forms the shape of a stylized roaring lion.

In contrast to older HIS cards, this one's fins are oriented vertically, unfortunately. Not only does this prevent heat from escaping from the slot bracket, but it also causes the motherboard to warm up. The top of the card is dominated by an illuminated HIS logo and eight-pin auxiliary power connector.

Around back, you get a peek at three of the four 6mm heat pipes. Meanwhile, the slot bracket exposes a familiar complement of outputs. Next to a cut-out styled to spell HIS, there's a DVI-I connector, three DisplayPort 1.4-ready interfaces, and one full-sized HDMI 2.0 port. Small vents run down the center, though they serve no practical purpose.

Board & Components

HIS puts the Ellesmere GPU on the board's left side, sticking with AMD's reference layout. As a result, the PCB's right side is sparsely populated with voltage regulation circuitry and a few other components. The company probably could have come up with a shorter design.

Like AMD, HIS uses eight Samsung GDDR5 memory modules (model number K4G80325FB-HC25). Each one has a capacity of 8Gb (32x 256Mb). And depending on clock rate, they can operate at voltages between 1.305V and 1.597V. Their ceiling is 2000 MHz, similar to what we find on Nvidia's GeForce GTX 1070.

The GPU is fed by six power phases using International Rectifier's IR3567B dual-output 6+2-phase controller. In turn, an IR CHL8510 high-performance gate driver controls the high- and low-side N-channel MOSFETs.

The actual voltage conversion is handled by an IRF6811 on the high side and an IRF6894 on the low side. The advantage of Infineon's DirectFET packaging is that it's easier to cool. However, the housing isn't isolated, so you need to keep an eye out for anything that might cause a short circuit.

HIS employs encapsulated ferrite coils covered by a thin-walled housing. Their quality is about average, as vibrations are both measurable and audible as a subtle buzzing noise.

The RX 480 IceQ X² Roaring Turbo relies on one phase to power its memory, located on the board's right side. A second, seemingly similar voltage converter provides power for peripherals and controllers. An Anpec APW8722 synchronous buck converter controls a Sinopower SM4377 on the high side and two SM4373s on the low side.

Power Results

This factory-overclocked card's power consumption is significantly higher than what we measured from the reference design. Its almost 20W at idle doesn't impress us, and we can see it's caused by rather generous voltage settings and a high frequency of 300 MHz.

During our gaming loop, the RX 480 IceQ X² Roaring Turbo consumes almost 200W, which is 30W higher than AMD's own solution. Meanwhile, our stress test pushed HIS' board as high as 215W.

We mention short load peaks only as a side note (see the light-grey bar in the chart below), since they're rarely relevant in practice. Nevertheless, they illustrate the importance of deploying a reasonably-sized secondary side for the power supply (low-impedance caps).

The following two graphs represent two minutes captured from the gaming loop and stress test. These are what our average power consumption figures are calculated from.

The next two graphs map current measurements at each of the supply rails, corresponding to the power consumption levels shown in the diagrams above.

Well, this isn't good. If you remember back to our Radeon RX 480 launch review, it was an out-of-spec load on the motherboard slot's +12V rail that forced AMD to add a compatibility mode to its driver. The values we measure for HIS' RX 480 IceQ X² Roaring Turbo are even worse, though!

A measurement of 7A exceeds the PCI-SIG spec's maximum of 5.5A by almost 30%! Even if modern boards can fall back on a fourth pin (meant as a reserve) to better distribute current, we definitely can't recommend this card to anyone looking to upgrade an older platform.

Temperature Results

HIS' cooler is both simple and conservative. It employs a backplate that is isolated with foil on the inside and does nothing to cool the board. Instead, the plate helps with stabilization and looks good.

The memory modules are not cooled by the heat sink, but rather by a small frame below it. You'd better believe we plan to quantify how well (or not) this concept works.

The cooler consists of a copper slug over the GPU and a total of four 6mm heat pipes pressed into it. Those pipes are made of a sintered composite material, and as you can see they aren't nickel-plated. A plate on the back connects the sink to one part of the finned structure. It's supported by the heat pipe that's bent back. Meanwhile, the other three pipes transport thermal energy the other direction, into another array of aluminum fins.

A peak clock rate of 1388 MHz during our gaming loop, which only drops by a few megahertz every once in a while, is indicative of the cooler's excellent performance. We blame the dips on a power limit that rears its head when we apply particularly taxing workloads.

Registering close to 152°F (67°C) on our open test bench and almost 158°F (70°C) in a closed case, the VRMs are even cooler than the GPU, which heats up to 158°F (70°C) and 163°F (73°C), respectively. The memory isn't particularly hot either, so we can confidently say that HIS does a fine job with cooling.

During the stress test, we measured 162°F (72°C) and 165°F (74°C) for the voltage converters. The processor is still fine as well, yielding a 163°F (73°C) result on our open test bench. Inside of a closed case, the PCB under the GPU's package reaches 169°F (76°C).

Again, this card's cooling is completely ample, even in a closed case.

Sound Results

The rotor geometry and blade count on the two 92mm fans are designed to provide high static pressure through the heat sink's dense array of fins. Thus, the fan speeds need to spin quickly in order to achieve the temperatures we measured.

A reading of roughly 3000 RPM in a closed case is high indeed (roughly 2500 RPM on an open test bench is up there, too). This naturally has an impact on the sound pressure level. During gaming, the almost 41 dB(A) we observed isn't particularly loud, but it's definitely audible. After all, the RX 480 IceQ X² Roaring Turbo has to dissipate about 200W of waste heat.

Low-frequency bearing noise at ~350 Hz dominates the spectrum chart, along with the fan itself between 2.5 and 4.5 kHz. We were also able to measure higher-frequency noises at approximately 16 KHz, corresponding to the voltage converters. Human ears are going to have a hard time hearing that, though.

HIS' RX 480 IceQ X² Roaring Turbo isn't a bad implementation of the Radeon RX 480. It does, however, suffer from an inappropriate distribution of power phases. This leads to current flow through the motherboard's PCIe slot that's almost 30% higher than the PCI-SIG's maximum specification.

This problem was exposed right when AMD launched the RX 480, so it's surprising that HIS not only ignores the issue but actually makes it worse, trampling upon specifications in the process.

The cooling performance we observed is quite convincing, and the fan noise is tolerable. An aggressive clock rate that's almost completely stable would have made this board worthy of recommendation if it wasn't for the overload condition. 

Thus, the way it sits currently, we cannot recommend the RX 480 IceQ X² Roaring Turbo. Instead, we have to suggest making sure your motherboard uses four pins, rather than the three standardized ones, for the card's power supply. If you don't know or can't tell, it's probably better to look for an alternative.

MORE: Nvidia GeForce GTX 1080 Roundup

MORE: Nvidia GeForce GTX 1070 Roundup

MORE: Nvidia GeForce GTX 1060 Roundup

MSI RX 480 Gaming X 8G

MSI leans on an old and trusted cooler for its RX 480 Gaming X 8G. A Gaming Mode clock rate of 1303 MHz isn't the most aggressive in our round-up, but it's certainly higher than AMD's reference design. Traditionally, MSI likes to implement an above-average power limit, a conservative fan control circuit, and a low temperature limit, so we're curious to see if this card behaves like so many Gaming X-series products before it.

Specifications

MORE: Best Graphics Cards

MORE: Desktop GPU Performance Hierarchy Table

MORE: All Graphics Content

Exterior & Interfaces

Weighing in at 974 grams, the RX 480 Gaming X 8G is fairly beefy. But it's not the heaviest card in our round-up. A length of 27.5cm, a height of 13.5cm, and a width of 3.5cm make this a large product; however, it still fits in a dual-slot form factor. Because of the backplate, you must plan for an additional one-fifth of an inch (5mm) of clearance to avoid collisions with your CPU's cooler, complicating installations in compact cases. 

You'd know this is an MSI card just by looking at it. Clearly, MSI's branding folks succeeded in establishing a recognizable corporate identity.

The plastic fan shroud is black with red highlights; it's attractive enough, though not necessarily the best-built in today's story. We already know the two fans well from MSI's other Gaming X products. The blades on the double-ball bearing fans have a diameter of 9.5cm, explaining the card's height.

Horizontally-oriented fins help guide hot air out of your case through the slot bracket and prevent heat from building up on the motherboard below the card. Up top, there's an illuminated MSI logo and an eight-pin power connector. We also catch a glimpse of two 6mm heat pipes that stretch from one end to the other, plus a shorter 8mm pipe.

The slot bracket exposes a familiar complement of outputs, including a DVI-I connector, two DisplayPort 1.4-ready interfaces, and two full-sized HDMI 2.0 ports (perhaps in a nod to the folks with VR HMDs). Several openings allow waste heat to pass through.

Board & Components

MSI positions the GPU's voltage regulation circuitry on the PCB's right side, while leaving two phases for the memory and peripherals on the left. This is opposite what we just saw from HIS.

Like AMD, MSI uses eight Samsung GDDR5 memory modules (model number K4G80325FB-HC25). Each one has a capacity of 8Gb (32x 256Mb). And depending on clock rate, they can operate at voltages between 1.305V and 1.597V. Their ceiling is 2000 MHz, similar to what we find on Nvidia's GeForce GTX 1070.

The GPU is fed by six power phases using International Rectifier's IR3567B dual-output 6+2-phase controller. In turn, an IR CHL8510 high-performance gate driver controls the high- and low-side N-channel MOSFETs.

MSI taps Ubiq Semiconductor's QM3816 dual N-channel MOSFET, which combines the high- and low-side transistors, as well as a Schottky diode on one chip. This saves space, but may also create an unnecessary hot spot if cooling isn't sufficient.

The RX 480 Gaming X 8G employs encapsulated ferrite coils covered by a thin-walled housing. MSI advertises these as Super Ferrite Chokes, and they're a bit classier-looking than what we've seen on competing cards.

One phase on the left side of MSI's board powers the memory. A second, seemingly similar voltage converter provides power for peripherals and controllers.

Power Results

This factory-overclocked card's power consumption is significantly higher than what we measured from the reference design. Its almost 18W at idle doesn't impress us, though we can see it's caused by rather generous voltage settings and a high frequency of 300 MHz.

During our gaming loop, the RX 480 Gaming X 8G consumes approximately 191W, which is significantly higher then AMD's own solution. Meanwhile, our stress test pushes MSI's board as high as 224W due to a fairly loose power limit.

We mention short load peaks only as a side note (see the light-grey bar in the chart below), since they're rarely relevant in practice. Nevertheless, they illustrate the importance of deploying a reasonably-sized secondary side for the power supply (low-impedance caps).

The following two graphs represent two minutes captured from the gaming loop and stress test. These are what our average power consumption figures are calculated from.

The next two graphs map current measurements at each of the supply rails, corresponding to the power consumption levels shown in the diagrams above. Unlike what we saw from HIS' card on the previous page, these charts should put your mind at ease.

A total of just 3.9A at the motherboard's PCIe slot is significantly lower than the 5.5A ceiling defined by the PCI-SIG. If you're worried about power consumption over the eight-pin auxiliary connector exceeding the recommended 150W, you shouldn't. During in-house tests, we successfully attempted to draw up to 400W of DC and at least 300W of AC loads over such a connection and saw no sign of overheating.

But because power consumption over the auxiliary power connector does exceed the specification by up to 22W, you should use a proper eight-pin plug, include 18-gauge wires. Six- or eight-pin adapters or the use of combined Molex connectors should be avoided!

Temperature Results

MSI's cooler employs a backplate that is isolated with foil on the inside and does nothing to cool the board. In addition, the card has a cooling frame that helps stabilize the backplate.

The memory modules are not cooled by the heat sink, but rather by the aforementioned frame installed below the larger thermal solution. MSI leaves the area around the coils and capacitors open so that they receive a bit of direct airflow.

The cooler itself consists of a nickel-plated copper sink for the GPU, one short 8mm heat pipe, and two longer 6mm heat pipes pressed into the base. Those pipes are made of sintered composite material and are nickel-plated as well.

MSI achieves solid thermal performance, evidenced by the fact that this card sustains its advertised 1303 MHz clock rate through our gaming loop.

The open-air and closed-case temperatures diverge during our stress test because of a power limit set too high. This causes inconsistency in the GPU frequency, too.

Registering about 176°F (80°C) during our gaming loop on an open test bench and inside of a closed case, MSI's cooling performance is acceptable. Interestingly, the GPU temperatures inside an enclosure is a little lower, even. Why? We'll get into that.

During our stress test, we measure 176°F (80°C) and 185°F (85°C) on the voltage converters. The processor fares well enough, peaking at around 163°F (73°C) on our open test bench. In the closed case, we see the GPU at 176°F (80°C). The warmest memory module remains below the 176°F (80°C) line as well.

So long as you don't make a habit of running stress tests, MSI's thermal solution gets the job done.

Sound Results

But what exactly causes the almost identical temperatures during our gaming loop on the open test bench and inside a closed case? One look at the fan speed logs yields an unpleasant surprise. Above a target temperature of 162°F (72°C), the fan curve is extremely aggressive to ensure the card gets no hotter. Unfortunately, the result is already clearly audible, and beyond what we are used to hearing from MSI.

Spinning at almost 2000 RPM, the RX 480 Gaming X 8G's fan is, installed in a closed case, one decibel louder than HIS' RX 480 IceQ X² Roaring Turbo (even if MSI's noise is a little more pleasant-sounding). We wonder why MSI doesn't at least allow the card to hit 176°F (80°C) before intervening in such an imposing way.

We also have to ding MSI for that peak between ~8 and 10 kHz. As you no doubt know, that's indicative of noisy voltage converter coils. Consequently, the slight buzzing/chirping sound is both measurable and audible.

In principle, MSI's RX 480 Gaming X 8G is a rock-solid, well-thought-out card that is missing some polish on its fan curve. A somewhat loose power limit leads to unnecessarily high power consumption under taxing workloads. This is made more bothersome by the fact it doesn't even translate to higher frame rates.

The cooling performance is adequate, but not good enough to allow for additional overclocking in a closed case without a bit of tweaking. You'd want to install strategically-placed fans to ensure constant airflow, which would naturally make the noise situation worse. Otherwise, expect to see high fan speeds uncharacteristic of MSI's usual efforts.

Definitely connect this card to your power supply with a native eight-pin connector using 18-gauge wires. Otherwise, leaning on adapters could prove detrimental to older or inexpensive PSUs.

MORE: Nvidia GeForce GTX 1080 Roundup

MORE: Nvidia GeForce GTX 1070 Roundup

MORE: Nvidia GeForce GTX 1060 Roundup

Sapphire Nitro+ Radeon RX 480

In comparison to Sapphire's simpler (and slightly newer) Nitro card, this one features elaborate RGB lighting and a second BIOS. You have to love that 1342 MHz maximum clock rate, too. But we wonder if it's sustainable under our gaming workload and stress test?

If you find the Nitro+ Radeon RX 480 too hot or loud by default, the second BIOS drops Sapphire's card down to a 1266 MHz boost frequency, similar to AMD's reference design. Because we test hardware the way it arrives out of the box, though, our numbers reflect the overclocked speed, heat, and noise. But don't worry, we'll comment on the "silent" BIOS when it's appropriate as well.

Specifications

MORE: Best Graphics Cards

MORE: Desktop GPU Performance Hierarchy Table

MORE: All Graphics Content

Exterior & Interfaces

Weighing just 774 grams, this is the lightest card in our test field, unfortunately hinting at a modest-sized heat sink. A length of 24.1cm makes the Nitro+ Radeon RX 480 pleasantly short. But that's countered by the rear-mounted auxiliary power connector. With a cable plugged in, the card is more like 27cm long.

Measuring 12.2cm tall and 3.5cm wide, it still fits in a dual-slot form factor at least. Because of the backplate, you must plan for an additional one-fifth of an inch (5mm) of clearance to avoid collisions with your CPU's cooler.

The plastic fan shroud is pretty conservative-looking. In contrast, the silver backplate decorated with decals and the Nitro logo is more of a stand-out. We know the two 95mm fans from Sapphire's other Nitro cards. Their blades have a net diameter of 90mm and are optimized for throughput rather than static pressure.

Horizontally-oriented fins help guide hot air out of your case through the slot bracket and prevent heat from building up on the motherboard below the card.

Up top, there's an illuminated Sapphire logo. The shroud wraps all the way around, hiding the heat sink underneath. We also spot the BIOS switch, which lets you toggle between overclocked and reference clock rates.

Around back, we catch a glimpse of the heat sink's fins behind an eight-pin power connector, while underneath we see two 8mm heat pipes and a smaller 6mm pipe.

Similar to MSI's complement of display outputs, Sapphire exposes one DVI-I connector, two DisplayPort 1.4-ready interfaces, and two full-sized HDMI 2.0 ports (perhaps in a nod to the folks with VR HMDs). Several openings allow waste heat to pass through.

Board & Components

Sapphire positions the GPU's voltage regulation circuitry on the PCB's right side, leaving one phase for the memory and peripherals on the left.

Like AMD, MSI uses eight Samsung GDDR5 memory modules (model number K4G80325FB-HC25). Each one has a capacity of 8Gb (32x 256Mb). And depending on clock rate, they can operate at voltages between 1.305V and 1.597V. Their ceiling is 2000 MHz, similar to what we find on Nvidia's GeForce GTX 1070.

The GPU is fed by five power phases using International Rectifier's IR3567B dual-output 6+2-phase controller (yes, this component could have allowed Sapphire to use six phases). Voltages are generated by one highly integrated IR3555 per phase, which combines the gate driver, high- and low-side MOSFETs, and Schottky diode in one package.

Although this conserves space on the PCB, it also creates hot-spots, as our measurements will substantiate. It wouldn't have been a bad idea to distribute that thermal load across multiple components.

The Nitro+ Radeon RX 480 employs encapsulated ferrite coils protected by a cover. These should be higher-quality than most of the chokes we find on reference cards.

In contrast, the memory's voltage regulation circuitry consists of just one phase. An Anpec APW8722 synchronous buck converter controls three ON Semiconductor NTMFS4C10N single N-channel MOSFETs (one is responsible for the high side and two take care of the low side).

A BIOS chip is located right under the small switch we saw earlier. It contains the default overclocked firmware and a more conservative reference fall-back.

Power Results

This factory-overclocked card's power consumption is higher than what we measured from the reference design—with an interesting twist. The 16W it uses at idle comes close to AMD's own solution, landing well below competing third-party products.

However, during our gaming loop, the Nitro+ Radeon RX 480 consumes approximately 223W. That's one of the highest observed values, and it's the result of aggressive clock rates, voltage settings, and a high power limit. Because the cooler isn't particularly beefy, though, the core frequency starts sliding backwards quickly under taxing loads, reducing power consumption in the process.

Finally, during the stress test, our power measurement drops to 119W, which is lower than the gaming loop.

At this point, it makes sense to talk about the fall-back BIOS as well, since it reduces power consumption close to the reference card's levels (16W at idle, ~165W while gaming, and 168W during our stress test). Of course, you pay a price in performance. But the resulting noise is much more pleasant.

The following two graphs represent two minutes captured from the gaming loop and stress test. These are what our average power consumption figures are calculated from.

The next two graphs map current measurements at each of the supply rails, corresponding to the power consumption levels shown in the diagrams above. Similar to what we saw from MSI's RX 480 Gaming X 8G, these charts should put your mind at ease.

A maximum of 5A (compared to the spec's 5.5A ceiling) shouldn't present any issues, even for older motherboards. If you're worried about power consumption over the eight-pin auxiliary connector exceeding the recommended 150W, you shouldn't. During in-house tests, we successfully attempted to draw up to 400W of DC and at least 300W of AC loads over such a connection and saw no sign of overheating.

But because power consumption over the auxiliary power connector does exceed the specification, you should use a proper eight-pin plug, include 18-gauge wires. Six- or eight-pin adapters or the use of combined Molex connectors should be avoided!

Temperature Results

Sapphire's thermal solution employs a backplate that is isolated with foil on the inside and designed to play a role in the card's cooling performance. It also helps add some rigidity.

Usually, Sapphire installs a thermal pad between the backplate and back of the board. It's positioned over the VRM, where it can facilitate heat transfer to the plate. In this case, however, apart from the fact that the backplate's insulation likely prevents any significant dissipation, the thermal pad is also crooked, warped, and misaligned. Superimposing the front and back of the card shows in red how far off the pad's placement is.

The memory modules are cooled directly by a frame attached to the heat sink. Voltage regulation circuitry is cooled by another sink built into the fin structure. Unfortunately, this heat sink doesn't help with the chokes, and we'll see that prove to be an oversight.

A copper sink over the GPU is mounted into the cooling frame. Turned over, this large mass is topped with one short 6mm heat pipe and two 8mm pipes. They aren't pressed into the sink, though.

The aforementioned fans employ fewer rotor blades than many competitors, but they're also a little larger. The resulting airflow is rather chaotic, which benefits the small cooler since static pressure doesn't help as much over limited surface area.

Unfortunately, Sapphire's cooling performance is merely mediocre. The card constantly stays below its 1342 MHz specification during our gaming loop, and the clock rate jumps around quite a bit. The reasons this card gets so hot are tied to its power limit (set way too high) and undersized cooler. Our stress test shows how the GPU frequency falls through the floor as the thermal threshold is reached.

Although switching over to the "silent" BIOS hits performance a bit, making it roughly equivalent to the reference design, it also reduces the Nitro+ Radeon RX 480's thermal load by up to 4 Kelvin on average.

Registering about 201°F (94°C) during our gaming loop on an open test bench, and 205°F (96°C) inside a closed case, this card is at the limit of what could be considered healthy. Sure, the MOSFETs can operate at up to 150°C without any damage. But over time, the PCB material can suffer irreversible damage, causing hairline cracks in the traces of multi-layer boards. At least the GPU and memory temperatures are still fairly safe.

During the stress test, we measure peaks of 203°F (95°C) and 207°F (97°C) for the voltage converters. The graphics processor is still in an acceptable range at 171°F (77°C) on our open test bench. In a closed case, it crests 174°F (79°C). Meanwhile, the warmest memory module stays below the 176°F (80°C) boundary.

Technically, the cooling is sufficient, even in a closed case, so long as you don't make a habit of running stress tests. But we have to say it seems like Sapphire got a little too aggressive with its default settings. The fall-back BIOS isn't as sexy, but it's more appropriate for the thermal solution Sapphire uses.

Sound Results

The fan speeds are harbingers of things to come. These curves are extremely aggressive, and we all know what that means...

Spinning at almost 2200 RPM, the fan is really loud in a closed case, yielding a measurement in excess of 47 dB(A). Compared to the other cards in our round-up, that's pretty high.

Not surprisingly, most of the noise is generated by the fans, which dominate the frequency spectrum. Possible bearing and voltage converter noises are drowned out and not individually discernible.

Using the secondary BIOS, we record a ~42 dB(A) reading. This is a big improvement, but it's still louder than the cards with big coolers.

Sapphire's Nitro+ Radeon RX 480 is not a bad card, per se. However, it suffers from a combination of overzealous factory overclocking, an undersized thermal solution, and relying on clearly audible airflow to compensate for a lack of cooling surface area. This is too bad, since the card's moderate dimensions would be a good fit for smaller cases.

The only workaround we can suggest is using the second BIOS. It's fortunate that Sapphire makes it easily accessible through a switch on top of the card. Operating at reference-class clock rates, the Nitro+ Radeon RX 480's fans are much quieter. You'll also want to connect this card to a power supply using a real eight-pin connector with 18-gauge cables, and not an adapter.

In the end, Sapphire's Nitro+ Radeon RX 480 is still worth considering. But expect it to be noisy, regardless of the BIOS you choose.

MORE: Nvidia GeForce GTX 1080 Roundup

MORE: Nvidia GeForce GTX 1070 Roundup

MORE: Nvidia GeForce GTX 1060 Roundup

Introduction & Overview

Nvidia launched its GeForce GTX 1060 6GB in July of 2016 to inevitable comparisons with AMD's Radeon RX 480 8GB. Although the 1060 was faster in DX11 games, it also commanded a premium that was harder to justify than the uncontested GeForce GTX 1070 and 1080.

A month later, Nvidia quietly rolled out a 3GB version of the 1060 to battle the 4GB RX 480. Its GPU took quite a haircut in the process, though, dropping from 1280 to 1152 CUDA cores and affecting performance far more than model name suggests. Presumably, Nvidia couldn't risk the 3GB and 6GB models appearing too similar at 1920x1080.

After wrapping up our initial GeForce GTX 1080 and GeForce GTX 1070 round-ups, Tom's Hardware DE set to work on a collection of 1060s, 3GB and 6GB alike. This first incarnation includes eight different boards from a field that spans anywhere from under $200 (£150) to over $300 (£250). Each individual review goes incredibly deep, covering manufacturing quality, technical features, power consumption, clock rates, cooling, and acoustics.

The gaming performance of every factory-overclocked board within a given chip class is usually pretty similar to begin with. But that's more true now than ever. This is a result of features like GPU Boost 3.0, which allows manufacturers to safely extract as much headroom as possible from a processor. Very little is left on the table, even if you have access to extreme overclocking hardware.

We will continue to update this roundup as new test samples become available.

The final analysis of each card is listed below for quick and easy comparison.

MORE: Best Graphics Cards

MORE: Desktop GPU Performance Hierarchy Table

MORE: All Graphics Content

MORE: Nvidia GeForce GTX 1080 Roundup

MORE: Nvidia GeForce GTX 1070 Roundup

MORE: All Graphics Content

Benchmark Results

We test every card after a suitable warm-up period to avoid unfair differences in GPU Boost frequencies. All benchmarks are run six times; the first one is used to get the GPU hot again.

These cards are all press samples operating at the same settings as retail models in our best effort to ensure one vendor doesn't get a leg up on another using non-representative clock rates.

The following galleries each contain four images, covering two tested resolutions. We put our focus on QHD (2560x1440) and FHD (1920x1080), plotting out average and minimum frame rates for each resolution in separate graphs.

Ashes of the Singularity

Battlefield 4

The Division

Grand Theft Auto V

Hitman (2016)

Project CARS

Rise of the Tomb Raider

The Witcher 3

All of the factory-overclocked cards offer similar performance, more or less. That's why our primary focus centers on evaluating the more technical aspects of each board design, along with their coolers. This is where differences in noise, power, and temperatures are most likely to come from.

Power Consumption

We begin by comparing the power consumption of each card in our gaming loop, stress test, and at idle. Depending on the power targets specified by each manufacturer, we sometimes measured substantial differences. We're ignoring the decimal places in our bar graphs, since those values are too small and would be within measurement tolerances.

We also found that some cards with lower power targets started to throttle during our stress test, resulting in lower power consumption numbers. This did not, however, have a negative impact on general gaming performance for any of the tested cards, as the stress test merely represents a worst-case scenario.

Noise

For the following comparison, we divide all of the gallery's bar graphs into gaming loop and idle, even if "noise" in practice spans a wide spectrum and the "character" of the sound varies a lot, with each card having highly individual results. Therefore, it is important not to compare just the absolute numbers, but also the frequency spectrum we're presenting.

Many of the cards implement a semi-passive mode, where their fans remain off when the card is idle. Thus, we refrained from taking measurements in that state. Even in our anechoic chamber, levels of 22 dB(A) and below merely represent ambient noise.

Temperature

For this comparison, we divide the gallery's bar graphs into gaming loop, stress test, and peak temperatures measured on the MOSFETs.

We occasionally compared the temperatures on our benchmark table with those measured inside a closed case and found them to be no more than two or three Kelvin higher.

Since temperatures in a closed case also depend heavily on the enclosure's cooling performance, the only representative and reproducible values are those measured on our benchmark table. Those are the ones we compare.

Nvidia GeForce GTX 1060 Founders Edition

For a reference design - sorry, Founders Edition - this card's appearance is quite neat. But are there any good reasons to pick Nvidia's implementation over competing models? After all, it's heavy and employs a completely different cooling solution than the competition.

Technical Specifications

MORE: Best Graphics Cards

MORE: Desktop GPU Performance Hierarchy Table

MORE: All Graphics Content

Exterior & Interfaces

Despite using a much shorter PCB, Nvidia's Founders Edition is still 10 inches (25.4cm) long. It is 4¼ inches (10.7cm) tall and 1½ inches (3.8cm) wide, too. The slot bracket is the widest part. Behind it, the card and cooler only measure about 3.5cm together.

For the shroud, Nvidia again uses a mix of aluminum and plastic. This time the company favors economy, though. The cover can be detached in one piece, including the fan.

There's still an illuminated GeForce GTX logo up top, and the cooler's internals are completely hidden.

In contrast to the coolers you often seen on cards with short PCBs, the housing's overhang doesn't have a hole to feed the fan with air. Instead, the back side is covered. There is no backplate either, further reducing cost.

The back hosts Nvidia's well-known, gill-shaped air intakes. These are more fragile, so you'll want to be careful positioning this card in a case. 

The output bracket carries over from Nvidia's GeForce GTX 1070 and 1080 Founders Editions. It's dominated by three DisplayPort 1.4-ready connectors, one HDMI 2.0 output, and one dual-link DVI-D interface.

Board & Components

The center of the card is naturally dominated by its graphics processor, even if the GP106-400-A1 is significantly smaller than the GP104 on GeForce GTX 1070 and 1080. There are other noticeable differences as well, like the on-board memory. Nvidia's GeForce GTX 1060 only uses six of the available eight module emplacements, each populated with Samsung K4G80325FB-HC25 ICs. The modules have a capacity of 8Gb (32x 256Mb), and operate between 1.305 and 1.597V, depending on clock rate. In total, they add up to 6GB of GDDR5.

Unfortunately, the PWM controller isn't documented. It’s made by uPI Semiconductor and bears the model number uP9509, which means that it’s probably the uP9511P’s smaller sibling (the latter controller is what we found paired to the GP104 processor).

The memory modules and one of the GPU phases get their power through the motherboard’s PCIe slot. The two remaining GPU phases and the card’s accessories draw power from the six-pin power connector. We'll take a closer look at what this means in terms of load distribution across the rails on the next page.

When it comes to voltage regulation, Nvidia uses only one Dual N-Channel MOSFET, the E6930, per phase for both the high and low side; separate gate drivers aren’t needed. This highly integrated component explains the empty spaces on the board.

The GPU’s three phases are completely sufficient, and their distribution makes more sense here than on AMD's Radeon RX 480.

Apart from the six-pin power connector, which appears to have taken a wrong turn somewhere, Nvidia's reference GeForce GTX 1060 actually looks pretty good. And given a relatively low amount of waste heat, its axial fan isn't a bad choice either.

Power Results

The GeForce GTX 1060 FE does hit its power and temperature limits.

In order to test overclocking, then, we not only increased the GPU Boost frequency, but also the power target to 116 percent while setting the fan to 100 percent. The results certainly speak for themselves, too. Our sample held a stable 2050 MHz through the most taxing passages of our test sequence. Without this forced bit of headroom and the additional fan speed, though, the graphs look shakier:

For comparison, let’s take a look at the voltages during normal operation as well. It’s plain to see that the GPU Boost clock and voltages drop once the limit is reached.

The lowest GPU frequency measured at idle was 139 MHz.

We took our measurements using a variable low-pass filter, so we mention short load peaks only as a side note (see the grayed-out bar in the chart below), since they're rarely relevant in practice. 

Power Connector Load

Now we're getting into the nitty gritty. Measurements across the different rails during our taxing gaming and stress tests show that Nvidia manages to distribute the load well with a bias to the six-pin power connector. The 3.3V rail isn't used anymore, which is why you won't find it in our charts.

Here are the detailed graphs. Clicking on them brings up the high-resolution versions.

The PCI-SIG’s technical specifications cover current exclusively, so that’s what’s shown in the chart below (power consumption only tells half of the story, after all). With less than 5A for the motherboard slot, Nvidia plays it safe with its GeForce GTX 1060. The maximum, according to the PCI-SIG, is 5.5A.

We have detailed graphs for the current measurements as well. Once again, clicking on them brings up the high-resolution versions.

Temperature Results

Up top, we immediately notice the power connector's strange position. It’s situated in a part of the cooler that protrudes beyond the actual PCA. This necessitates a number of cables to attach to the board.

The implementation is anything but elegant, and it prevents Nvidia's partners from building shorter 1060s. Although the card is only 17.5cm long, it doesn't have any space to accommodate a power connector.

Remove the four screws securing the cooler's body and it comes right off. There’s a massive copper heat sink and metal frame underneath. The closed cooling fin design reminds us of the GeForce GTX 1070, and it should provide ample performance given the 1060's 120W TDP.

The massive frame serves double duty by keeping everything in place and cooling the voltage regulation circuitry/memory modules.

The GeForce GTX 1060 generates similar numbers during the gaming loop (119W) and stress test (122W). In light of this, it’s hardly surprising that the two temperature curves look a lot alike. GPU Boost 3.0 is very restrictive (or already well-optimized, if you prefer). It adjusts the clock rate and voltage to keep the card exactly where it’s supposed to be.

The infrared picture tells us that the cooling solution does its job well. All four of the voltage converter areas are in great shape as well.

During the stress test, the GPU’s three voltage converters reach 97 degrees Celsius, which is just within the acceptable range. Then again, nobody runs stress tests for hours, so this result is more theoretical, and not so practical, in nature.

Sound Results

Next, we want to quantify just how loud the GeForce GTX 1060 gets under these conditions. The largest source of noise is generally the fan, so we look at that first. Its rotational speed takes a while to plateau, finally stabilizing after about 15 minutes. The two curves are almost on top of each other, which is again hardly surprising seeing that the temperatures the cooler deals with are almost identical as well.

At idle, we measure 31.4 dB(A). This is a good result that’s barely above the noise level encountered in a living room. The sound produced by Nvidia’s radial fan is noticeably more pleasant than that of AMD's reference Radeon RX 480. Nvidia's fan is a bit louder, though.

The good news continues with our gaming loop results. After reaching its maximum temperature, the GeForce GTX 1060 emits just 35.4 dB(A). Motor and bearing noises can’t really be heard above the relatively pleasant swooshing noise produced by the card. The sound isn't as low and growling as what we hear from AMD's competing offering. For more on that card, check out AMD Radeon RX 480 8GB Review.

Even a casual glance at this graph reveals that the majority of the GeForce GTX 1060’s operating noise comes from the fan blades and airflow. The voltage regulation circuitry doesn't contribute at all. This means that the DHE cooling solution performs well, doing its job quietly. Then again, it's not that big of a challenge to dissipate 120W.

MORE: Nvidia GeForce GTX 1080 Roundup

MORE: Nvidia GeForce GTX 1070 Roundup

MORE: All Graphics Content

EVGA GeForce GTX 1060 SC Gaming

With the 6GB GeForce GTX 1060 SC Gaming, EVGA added a model that roughly follows Nvidia's reference design, but uses a cooler that eliminates the Founders Edition's overhangs. This gives us a much shorter card seemingly ideal for HTPCs.

This model has some minor weaknesses, though. We've already discussed them with EVGA, and the company apparently plans to update its BIOS after checking it with Nvidia. Improved VRM cooling is on the table as well. However, since EVGA allows customers to change its cooler, you can apply our suggested modifications in a few manual steps without voiding the card's warranty.

Technical Specifications

MORE: Best Graphics Cards

MORE: Desktop GPU Performance Hierarchy Table

MORE: All Graphics Content

Exterior & Interfaces

The cooler cover is a low-cost affair made of anthracite-gray plastic and devoid of gaudy extras. Perhaps that's a plus. In compact cases, the card probably won't be visible anyways. Functional and affordable are interesting attributes in their own right.

Measuring just 7¼ inches (18.4cm) long from the slot cover, 4 1/5 inches (10.5cm) tall from the top edge of the motherboard slot, 1 3/8 inches deep, and weighing just 20 ounces (576g), the GeForce GTX 1060 SC Gaming comes across as a light-weight dwarf. And if space if your chassis is still an issue, it's easy to cut another third of an inch off the shroud if you have the right tools.

EVGA also abstained from attaching a backplate for reasons of cost efficiency, which frees up roughly one-fifth of an inch (5mm) for the CPU cooler in a mini-ITX enclosure. Sometimes that's the difference between success and failure in a compact build.

There's an EVGA logo up top, but it isn't back-lit. Moreover, a simple six-pin power connector sits at the end of the card and is not rotated.

A closer look at the bottom shows that, in spite of the card's compact design, EVGA employed a heat pipe-based cooling solution anyway. Our upcoming measurements speak to the wisdom of that decision.

Horizontally-aligned fins help some of the GPU's waste heat escape out the slot cover, which is fairly free-flowing thanks to a number of holes cut into it. This principle is also apparent at the end of the card, where hot air dumps out into the middle of your case. Good circulation is critical as a result.

The slot cover exposes five outputs, four of which can be used simultaneously to drive multi-monitor configurations. In addition to one dual-link DVI output (which lacks an analog signal), you also get one HDMI 2.0 interface and three DisplayPort 1.4-capable connectors.

Board & Components

EVGA's PCB is similar to Nvidia's reference design, but still unique in detail. For instance, while the GeForce GTX 1060 Founders Edition is manufactured in China, EVGA relies on a Taiwanese supplier.

Right off the bat we see four power phases, just like Nvidia's card. Three are meant for the GPU and one corresponds to the memory. Two of the three GPU phases are fed by the auxiliary power connector. The third GPU phase and the one for the memory get power from the motherboard slot.

In theory, the PCB could have enabled three phases and a maximum of six regulator circuits via doubling for the GPU, plus one phase for the memory. However, this minimalist configuration with just one regulator circuit per phase, similar to Nvidia's reference design, seems to be more popular given its lower cost.

The three GPU phases are controlled using one uPI Semiconductor uP9505P, with each regulator circuit realized as a highly integrated E6930 Dual N-Channel MOSFET, which combines the gate driver, high- and low-side MOSFET, and Schottky-diode in one convenient package.

Again, this is meant to cut costs and conserve space. But concentrating all of the components in one package also comes with a significant disadvantage: the creation of thermal hot-spots. This is made worse by the fact that two of these MOSFETS are positioned fairly close to each other. Since the PCB was designed to facilitate doubling, one of the two MOSFETS could have been moved into the second regulator circuit's vacant space.

Two capacitors are installed right below the GPU to absorb and equalize voltage peaks.

Only six of the eight available memory emplacements are populated with Samsung K4G80325FB-HC25 modules. Each one adds 8Gb (32x256Mb) and operates at voltages between 1.305 and 1.597V, depending on clock rate. In total, they add up to the GeForce GTX 1060 SC Gaming's 6GB of GDDR5.

Power Results

A compact graphics card typically enjoys less thermal headroom due to its smaller cooler. Naturally, we want to track the correlation between clock rate and voltage.

The data shows us that frequency drops slightly as the card heats up during a gaming session.

The 2012 MHz GPU Boost frequency we initially measure falls to 1949-1962 MHz under load. Voltage behaves similarly: after first measuring 1.062V, that value dropped to 0.975V on average after warming up (the actual range was 1.025 to 0.9V). All of this points back to GPU Boost 3.0 and a relatively low power target that limits consumption to just over 120W.

The lowest GPU clock rate measured with an idle card was about 253 MHz, and at that frequency, power consumption was similar to Nvidia's reference design.

We took our measurements using a variable low-pass filter, so we mention short load peaks only as a side note (see the grayed-out bar in the chart below), since those peaks are rarely relevant in practice. Nevertheless, that value is quite high, hinting at less than optimal primary-side smoothing prior to the voltage regulators.

A low power target causes GPU Boost to cut the voltages dramatically during our stress test, resulting in power consumption numbers that are lower than what we measure during a worst-case gaming workload.

Power Connector Load

Load distribution from the four voltage regulators to their respective power phases isn't problematic at all. However, there's little headroom for overclocking, even if it wasn't already limited by the firmware's power target and voltage settings.

Here's the data for the gaming and stress tests:

Power consumption on its own only tells us part of the story. The PCI-SIG specifies a maximum of 5.5A at the motherboard slot, and we're showing a reading under 5A, which puts EVGA's GeForce GTX 1060 SC Gaming in the green.

Of course, we have the data for measured currents as well:

The Cooler, Its Performance & A Simple Mod

Fortunately, the headaches caused by EVGA's cooler can be improved in just a few steps. Since the company does permit removal of its thermal solution without voiding the card's warranty, we can describe our modification in detail and encourage you to try it yourself. As long as EVGA doesn't implement changes of its own to augment performance, this procedure is worthwhile.

Once the four screws holding the cooler above the GPU package are removed and the cooler lifted off, two flat 8mm heat pipes become visible. They're meant to dissipate heat from the copper sink to the fins on both ends of the cooler. Their composite material should be able to deal with curvature and orientation in a reasonably efficient manner.

However, it's apparent that the VRM's four dual-channel MOSFETs and the memory packages receive no active cooling. This is exactly where we will apply our minor modification to affect big change with little effort.

The built-in 9cm fan relies on conventional rotor blades, which are quite steep to generate more pressure. Unfortunately, the heat pipes cover several important areas, such as the VRMs, meaning little of that extra airflow actually reaches the PCB's surface.

As far as the GPU is concerned, we measure 163 to 165°F (73 to 74°C) during our gaming loop and a maximum of 172°F (78°C) in a closed case. When it comes to the processor itself, EVGA's cooler does a respectable job. The stress test reveals a more gradual ramp up due to a lower frequency and less power consumption. But the final temperatures are nearly identical.

As our infrared readings demonstrate, though, a lack of active cooling on the VRMs has negative consequences. Even on an open test bench, temperatures during our gaming loop reach 234°F (112°C). Our stress test gets them up to 246°F (119°C). Although that sounds bad, it's not out of spec for the built-in MOSFETs, which are rated at up to 302°F (150°C). However, the excess heat travels through the multi-layer circuit board's copper traces and spreads over a large area.

The PWM controller mounted on the back of the board (visible just right of the VRM label as a faint rectangle) measures approximately 221°F (105°C). Again, this is within an acceptable range, since only temperatures as high as 302°F (150°C) would end in an OTP-related shut-down. We're more worried about the capacitors. Their through-hole leads work like little antennas, effectively transporting heat inside. With well over 212°F (100°C), and up to 234°F (112°C) during the stress test, this could lead to problems over time.

After all, these measurements were taken in an air-conditioned room at 72°F (22°C). In the height of summer, smaller PC cases heat up considerably. The same readings collected under less forgiving circumstances would likely turn out to be even higher.

Furthermore, the fact that the circuit board just below the GPU package is hotter than what the processor's own diode measures means the GPU is cooling the PCB, which is counterproductive. We do hope EVGA considers implementing something similar to our remedy.

A Safety Modification

The fact that the heat pipes cover the MOSFETs can easily be used as leverage. For the sake of simplicity, this example focuses exclusively on reducing the larger hot-spot of the two upper GPU VRMs by applying a stack of two thermal pads to bridge the tenth of an inch (2.5mm) difference between them and the heat pipe. Of course, you'd want to set up all four MOSFETs like this, yielding an even better result.

This superimposed image clearly shows the position of the heat pipes in relation to the components underneath, along with the position of our small cooling aid.

Retaking the measurement with an infrared camera left us amazed. Despite our meager investment and basic materials, the impact on cooling was astonishing. A reading seven to eight Kelvin lower for only two of the four MOSFETs puts us in a temperature range that allows some confidence in long-term reliability, even in small cases.

Sound Results

Hysteresis is not an issue for most manufacturers, and EVGA typically fares well here. After applying the most recent BIOS update, the GeForce GTX 1060 SC Gaming's semi-passive mode worked well without any annoying start/stop cycles. The threshold at which the fan spins up is set to ~149°F (65°C), but the GPU's current load isn't taken into account.

It's possible to lower the GPU's temperature four or five Kelvin more by configuring a custom fan curve to give the card a bit of extra air (important for mini-ITX projects). That helps the GPU Boost clock rates, too.

When the card is idle, noise is not measurable due to a semi-passive mode. A reading of 34.3 dB(A) under load is commendable, and that's primarily due to a conservative fan curve. However, even at 1300 RPM the noise level remains below the still-acceptable 36 dB(A) mark.

MORE: Nvidia GeForce GTX 1080 Roundup

MORE: Nvidia GeForce GTX 1070 Roundup

MORE: All Graphics Content

Gainward GeForce GTX 1060 6GB Phoenix GS

With its GeForce GTX 1060 6GB Phoenix "GS", Gainward aims for a slightly more demanding enthusiast without giving in to excessive overclocking madness. You can think of the company as a sort of luxury brand belonging to Palit, so it benefits from the organization's substantial resources. It's only a shame that our U.S. readers don't have access to the board. According to Gainward's webpage, the GeForce GTX 1060 6GB Phoenix "GS" is only available in Europe and Asia.

Gainward labels its top-performing cards in a given product family with the Golden Sample (GS) suffix, and that's exactly what our German lab received for testing.

Technical Specifications

MORE: Best Graphics Cards

MORE: Desktop GPU Performance Hierarchy Table

MORE: All Graphics Content

Exterior & Interfaces

The cooler cover is made of black plastic with various metal highlights. It is designed to be used as an attachment, just like the fan modules. Thus, a Palit card could easily be transformed into a Gainward one and vice versa.

Although the card weighs a moderate 31oz (875 grams), it is fairly compact, measuring 9½ inches (24cm) long and 4 3/5 inches (11.7cm) tall. However, it's also 4.7cm wide, nearly requiring three expansion slots and making installation in mini-ITX motherboards difficult.

The back of the board is covered by a single-piece plate that lacks openings for ventilation and does nothing to cool the PCB. Because of the backplate, though, you must plan for an additional one-fifth of an inch (5mm) of clearance to avoid collisions with your CPU's cooler, particularly in small form factor cases.

The top of the card features a Gainward logo back-lit with customizable RGB LEDs. A single six-pin power connector is rotated by 180° and positioned at the end of the card.

A closer look at the bottom reveals that the heat pipes aren't embedded in a heat sink, but rather lay flat behind the base of the board. You can bet we'll explore how well this cost-effective implementation holds up to taxing graphics workloads.

Horizontally-oriented fins allow some waste heat to escape from an externally-facing slot bracket loaded with ventilation holes. This also applies to the card's other end, though, where hot air dumps out into the middle of your case. Good circulation is critical as a result.

The slot cover exposes five outputs, four of which can be used simultaneously to drive multi-monitor configurations. In addition to one dual-link DVI output (which lacks an analog signal), you also get one HDMI 2.0 interface and three DisplayPort 1.4-capable connectors.

Board & Components

The PCB that Palit and Gainward use features some prominent differences compared to Nvidia's reference design.

Right off the bat we see four power phases, just like Nvidia's card. Three are meant for the GPU and one corresponds to the memory. Two of the three GPU phases are fed by the auxiliary power connector. The third GPU phase and the one for the memory get power from the motherboard slot.

While, in theory, Gainward's PCB offers the option to use one more power phase, the company went with a rarely-used and rather inexpensive NCP81174 for its PWM controller, which can only control up to four phases. Thus, the vacant emplacement you see in the image above could only have been populated by another parallel regulator circuit (via doubling).

The three GPU phases utilize a highly integrated Vishay SiC632, which combines the gate driver, high- and low-side MOSFET, and Schottky-diode in one convenient package. Meanwhile, the memory is driven by a pair of N-Channel MOSFETs, including one SM4503NHKP (30A, high-side) and one SM4377NSKP (50A, low-side), both manufactured by Sinopower. The MOSFETs are also fairly cost-effective choices.

Overall, the board reflects that Gainward didn't spend big on pricey components. Even the Foxconn Magic chokes are middle-of-the-road. They get their job done in a rather unremarkable way.

Two capacitors are installed right below the GPU to absorb and equalize voltage peaks.

Only six of the eight available memory emplacements are populated with Samsung K4G80325FB-HC25 modules. Each one adds 8Gb (32x256Mb) and operates at voltages between 1.305 and 1.597V, depending on clock rate. In total, they add up to this card's 6GB of GDDR5.

Power Results

The GPU Boost frequency and core voltage curves are particularly interesting to us. Our data shows that frequency drops slightly as the card heats up during a gaming session.

The 2025 MHz GPU Boost frequency we initially measure falls to 1949 MHz under load. Voltage behaves similarly: we measured 0.975V on average, with readings varying between 1.025 and 0.9V. All of this points back to GPU Boost 3.0 and a relatively low power target that limits consumption to a maximum of 130W.

The lowest GPU clock rate measured with an idle card is about 253 MHz.

We took our measurements using a variable low-pass filter, so we mention short load peaks only as a side note (see the grayed-out bar in the chart below), since those peaks are rarely relevant in practice.

A low power target causes GPU Boost to cut the voltages dramatically during our stress test, resulting in power consumption numbers that are lower than what we measure during a worst-case gaming workload.

Power Connector Load

Load distribution from the four voltage regulators to their respective power phases isn't problematic at all. However, there's little headroom for overclocking, even if it wasn't already limited by the firmware's power target and voltage settings.

Here's the data for the gaming and stress tests:

Power consumption on its own only tells us part of the story. The PCI-SIG specifies a maximum of 5.5A at the motherboard slot, and we're showing a reading under 5A, which puts Gainward's GeForce GTX 1060 6GB Phoenix "GS" in the green.

Of course, we have the data for measured currents as well:

The Cooler And Its Performance

As mentioned, Gainward's backplate is purely aesthetic; it doesn't help cool the PCB at all. Four screws secure it to the top of the board, shielding the card from curious customers.

Though the Gainward and Palit shrouds look a little different, everything else is mostly identical (at least up to the geometry of their respective 9cm fans; Gainward went with a conventional straight edge).

Again, the delicate-looking cooler sports horizontally-oriented fins, as well as a quartet of 4mm heat pipes made from sintered composite material. Those pipes are not integrated into the bottom of the cooler. Instead, they merely touch the back of the thin base plate. Expensive copper is nowhere to be found.

We're glad to see that the base plate has a specific area to help cool the MOSFETs. However, the coils have to live without the benefit of active cooling.

Registering between 156 and 158°F (69 to 70°C) in our gaming loop, and a maximum of 167°F (75°C) in a closed case, the cooler does a fair job. It would almost assuredly perform better if the heat pipes were attached more securely, though.

Measurements collected with the backplate removed also reveal acceptable values. A peak of approximately 181°F (83°C) from the VRMs is reasonable, in our opinion.

While the readings generated during our stress test aren't exactly worrisome at ~192°F (89°C), they also reveal the cooling solution's weaknesses.

Gainward's cooling performance is thus regarded as merely acceptable. We get the impression that one member of the finance team was assigned to each engineer, keeping costs in check more than enthusiasts like to see.

Sound Results

Hysteresis is not an issue for most manufacturers. However, Gainward implemented a fan curve with too-low of an initial rotational speed for the transition between passive and active mode. As a result, the fans simply stop, and need to be restarted repeatedly. Even the best hysteresis won't help when a manufacturer overestimates the quality of its components this grossly.

When the card is idle, noise is not measurable due to a semi-passive mode. A reading of 35.6 dB(A) under load is acceptable, but not great. So much air has to be moved through the thermal solution because Gainward's cooler isn't as efficient as it should be. Spending a few dollars more would have gone a long way. On the bright side, we measure very little low-frequency bearing noise. What remains is the sort of acoustic energy that's easily absorbed by a case with sound insulation.

MORE: Nvidia GeForce GTX 1080 Roundup

MORE: Nvidia GeForce GTX 1070 Roundup

MORE: All Graphics Content

Gigabyte GeForce GTX 1060 G1 Gaming 6G

Gigabyte's GeForce GTX 1060 G1 Gaming 6G is not the manufacturer's least-expensive version. Rather, it targets price-conscious gamers looking for a good deal, but still interested in feature-rich hardware. Thus, this factory-overclocked card represents the mass-market model of Gigabyte's GeForce GTX 1060s. The company reserves shorter ITX-oriented models for its system integrator business and doesn't aggressively advertise them directly to enthusiasts. Instead, the GeForce GTX 1060 G1 Gaming 6G employs a full-size PCB covered by a relatively large cooler.

Technical Specifications

MORE: Best Graphics Cards

MORE: Desktop GPU Performance Hierarchy Table

MORE: All Graphics Content

Exterior & Interfaces

The fan shroud is made of black plastic and seems fairly inconspicuous, despite some orange detail work. The two 3½-inch (9cm) fans are well known from previous Windforce-equipped models, and (from an aesthetic perspective) seem a bit lost on this otherwise long card.

Weighing in at just under 29 ounces (813g), this card is light for its size, particularly given measurements of more than 11 inches (28cm) long, 4 1/3 inches (11cm) tall, and 1 1/3 inches (3.5cm) wide. It occupies two full expansion slots, and is neither particularly small nor really oversized.

We're glad that Gigabyte decided against the almost obligatory third fan present on many of its coolers. Frankly, another one would have been completely superfluous for this card. A little more distance between the fans might have worked even better, though.

The card is covered by a single-piece backplate and lacks openings for ventilation. But it is connected to the board in two spots via thermal tape to aid active cooling. Because of the backplate, though, you must plan for an additional one-fifth of an inch (5mm) of clearance to avoid collisions with your CPU's cooler, particularly in small form factor cases.

Up top, there's a back-lit Gigabyte logo and one indicator for the semi-passive mode. Both are enabled as customizable RGB LEDs. An eight-pin auxiliary power connector is rotated 180° and positioned at the end of the card.

Vertically-oriented cooling fins mean that hardly any air escapes through the slot bracket, even though narrow slits in the metal suggest otherwise. The vertical fins are cheaper to implement, since they make it easier to arrange the heat pipes. You just have to be diligent about exhausting all of that waste heat out of your case.

The slot cover exposes five outputs, four of which can be used simultaneously to drive multi-monitor configurations. In addition to one dual-link DVI output (which lacks an analog signal), you also get one HDMI 2.0 interface and three DisplayPort 1.4-capable connectors.

Board & Components

A look at the board reveals that it was inspired by Gigabyte's GeForce GTX 1080 G1 Gaming and GeForce GTX 1070 G1 Gaming, but with far fewer components since the 1060 is electrically less complex.

Gigabyte relies on a rather unusual 6+1-phase design, with a uPI Semiconductor uP9511P PWM controller for the GPU's six phases, leaving two emplacements vacant. The memory's power comes from a seventh phase controlled by a uP1666; it consists of a single regulator circuit.

Since the PWM controller can't communicate directly with each phase's MOSFETs, Gigabyte utilizes separate PWM drivers (gate drivers) to control the two single-channel MOSFETs per phase for the range of the DC/DC voltage converter (low- and high-side). The second low-side MOSFET, which the higher-end cards do have, is missing in this case.

Two capacitors are installed right below the GPU to absorb and equalize voltage peaks.

Only six of the eight available memory emplacements are populated with Samsung K4G80325FB-HC25 modules. Each one adds 8Gb (32x256Mb) and operates at voltages between 1.305 and 1.597V, depending on clock rate. In total, they add up to this card's 6GB of GDDR5.

Power Results

Before addressing power consumption directly, let's take a look at the relationship between clock rate and voltage. As we've seen already, the GPU's frequency drops significantly once the card is fully exercised with a gaming workload.

After warm-up, the initial GPU Boost clock rate of 2000 MHz falls to a still-respectable 1974 MHz under load. The slope of this curve is mirrored by our voltage measurements, which aren't much worse either (on average, we measured 1.05V). This value drops slightly, but only sporadically.

The lowest GPU clock rate measured at idle is about 240 MHz.

We took our measurements using a variable low-pass filter, so we mention short load peaks only as a side note (see the grayed-out bar in the chart below), since those peaks are rarely relevant in practice.

Power Connector Load

Since Gigabyte's GeForce GTX 1060 G1 Gaming doesn't use the motherboard slot's 3.3V rail, we left it out of our chart.

The power distribution we observe shows a heavy bias toward the auxiliary power connector, which supplies about 100W.

Here's the data for the gaming and stress tests:

Power consumption on its own only tells us part of the story. The PCI-SIG specifies a maximum of 5.5A at the motherboard slot, and we're showing a reading under 3A, which puts Gigabyte's GeForce GTX 1060 G1 Gaming 6G in the green.

Of course, we have the data for measured currents as well:

The Cooler & Its Performance

When Gigabyte first launched this card, it didn't have thermal pads on the backplate to help draw heat from the board. This changed during mass production, however, after we passed along our feedback.

The revised version wasn't available at the time this review was originally written, so we had to implement the changes Gigabyte made ourselves. Similar to the GeForce GTX 1080 G1 Gaming, there are now two pads on the plate. They're in slightly different places, though, since a 1060's hot-spots aren't quite the same. One corresponds to the VRMs, while the other sits opposite Nvidia's GPU.

We started by taking readings with the card as it shipped from Gigabyte, and then introduced our modifications, which we can judge based on their effect on thermal performance.

The cooler employs vertically-oriented fins and two angled 6mm heat pipes made of a sintered composite material. They're flattened to sit flush with the heat sink, and directly touch the GPU. For cost reasons, the sink is all-aluminum, while the supporting base plate is fashioned from steel to improve stability.

It deserves a positive mention that the heat sink for the voltage regulation circuitry is built into the cooler and does an excellent job.

A measurement between 152 and 154°F (67 to 68°C) during our gaming loop, peaking at 158°F (70°C) in a closed case, proves that the cooler does an acceptable job. Based on our do-it-yourself attempt at replicating Gigabyte's modified version, we'd expect to see temperatures about two Kelvin lower.

Measurements taken with the backplate removed show a maximum of ~160°F (71°C) around the VRMs. That's an excellent value for what often appears as a troubling hot-spot on competing cards. Thermal pads should help even more, too.

In both our gaming workload and stress test, Gigabyte's cooler works the way it should, yielding satisfactory infrared images.

Although we were initially skeptical of how well the two flattened heat pipes would fare, the thermal solution performs as well as we could ask. Two 8mm or three 6mm heat pipes would have made more sense to us since they would have covered the GPU completely. But it's good to see Gigabyte cooling the front and back of its card in such a way that hot-spots are minimized.

Sound Results

Implementing exemplary hysteresis is a matter of honor for Gigabyte, and the company doesn't disappoint. With that said, a fan speed just above 1400 RPM shows where this solution's performance comes from. Of course, you can always play with the fan curve yourself. Before we do that, let's see what the stock noise levels sounds like.

When the card is idle, its noise level is imperceptible due to a semi-passive mode. There's no point taking readings in that state.

Under load, a 34 dB(A) reading is commendable, even if it isn't completely silent. In a closed case, the GeForce GTX 1060 G1 Gaming would only be audible as a soft whisper. The bearing noises we measure are acceptable, and there is very little unwanted structure-borne noise.

Relatively aggressive factory overclocking turns this fairly quiet card into a capable performer. It doesn't slip up in any of the disciplines we consider important, and lands at the top of our results in each individual test. A high GPU Boost frequency, which remains stable even after we warm the card up, make this an interesting choice for additional overclocking.

MORE: Nvidia GeForce GTX 1080 Roundup

MORE: Nvidia GeForce GTX 1070 Roundup

MORE: All Graphics Content

Inno3D GeForce GTX 1060 Gaming OC

Inno3D's GeForce GTX 1060 Gaming OC isn't the company's flagship, but rather a bread-and-butter workhorse in its 11-card portfolio of GP106-based products. We consider this a good thing. As one of the more affordable GeForce GTX 1060s (at least in Europe), the Gaming OC board appeals to a broader audience. You get a similar level of performance without spending an extra $40 or $50 (£20-30) on bells and whistles you might not need. It'll be interesting to see where Inno3D cut costs though, and what features it dropped to achieve a more attractive price.

Technical Specifications

MORE: Best Graphics Cards

MORE: Desktop GPU Performance Hierarchy Table

MORE: All Graphics Content

Exterior & Interfaces

The fan shroud is made of black plastic and features red detailing that looks quite nice. We won't say the same for those piano-black bolt-ons that serve little purpose other than to collect fingerprints. After installing the card, you'll feel compelled to clean it off again.

Weighing in at 29 ounces (820g), this card is fairly light. It measures 11 inches (28cm) long, five inches (12.5cm) tall, and 1 1/3 inches (3.5cm) wide. It monopolizes two expansion slots and is neither particularly small nor oversized—it's just average. The blades of both fans have a diameter of 9cm, which is pretty common as well.

The back of the board is covered by a single-piece plate with some ventilation holes. Unfortunately, there are no thermal pads connecting it to the circuit board, so the plate does nothing to help with cooling. Because of the backplate, though, you must plan for an additional one-fifth of an inch (5mm) of clearance to avoid collisions with your CPU's cooler, particularly in small form factor cases. A lack of SLI support means you don't have to worry about fitting two GeForce GTX 1060s back to back in a case.

The top of the card features an unlit Inno3D label, and the shroud up there is narrow enough that you see the cooling fins behind it.

An eight-pin auxiliary power connector is located at the end of the card facing out, rather than back. As a result, the heat sink is cut out right above the connector. Interestingly, this can be explained by the fact that Inno3D uses Nvidia's reference PCB for the GeForce GTX 1070 to build its GTX 1060 Gaming OC.

Vertically-oriented cooling fins mean that hardly any air escapes through the slot bracket, even though narrow slits in the metal suggest otherwise. The vertical fins are cheaper to implement, since they make it easier to arrange the heat pipes. You just have to be diligent about exhausting all of that waste heat out of your case.

The slot cover exposes five outputs, four of which can be used simultaneously to drive multi-monitor configurations. In addition to one dual-link DVI output (which lacks an analog signal), you also get one HDMI 2.0 interface and three DisplayPort 1.4-capable connectors.

Board & Components

A glance at the board gives us a sense of déjà vu; as mentioned, Inno3D relies on the longer GeForce GTX 1070 reference board. Of course, the components it uses to build the 1060 are simpler, but because Nvidia's GP106 processor is pin-compatible, it's not a problem to use the 1070's PCB.

The 5+1-phase design, similar to Nvidia's GeForce GTX 1070 Founders Edition, relies on the uPI Semiconductor uP9511P as its PWM controller. All five of the GPU's power phases are fed by the eight-pin auxiliary connector; only the memory's phase is driven through the motherboard's PCIe slot.

The voltage of the five GPU phases and one memory phase is controlled by no-name dual-N-channel MOSFETs that require external gate drivers. Inno3D's coils of choice are clones of Foxconn's Magic series, which we'd expect to find in the lower price segment.

Two capacitors are installed right below the GPU to absorb and equalize voltage peaks.

It's worth noting that only six of the eight available memory emplacements are populated with Samsung K4G80325FB-HC25 modules. Each one adds 8Gb (32x256Mb) and operates at voltages between 1.305 and 1.597V, depending on clock rate. In total, they add up to this card's 6GB of GDDR5.

Power Results

Before addressing power consumption directly, let's take a look at the relationship between clock rate and voltage. As we've seen already, the GPU's frequency drops significantly once the card is fully exercised with a gaming workload.

After warm-up, the initial GPU Boost frequency of 1974 MHz falls to an acceptable average of 1949 MHz under load. The slope of this curve is mirrored by our voltage measurements: what started at 1.043V in the beginning falls to an average of 1V. The drops we observe correspond closely with dips in GPU clock rates.

The lowest GPU clock rate measured at idle is about 203 MHz.

We took our measurements using a variable low-pass filter, so we mention short load peaks only as a side note (see the grayed-out bar in the chart below), since those peaks are rarely relevant in practice.

Power Connector Load

Since Inno3D's GeForce GTX 1060 Gaming OC doesn't use the motherboard slot's 3.3V rail, we left it out of our chart.

The power distribution we observe shows a heavy bias toward the auxiliary power connector, which supplies about 95W.

Here's the data for the gaming and stress tests:

Power consumption on its own only tells us part of the story. The PCI-SIG specifies a maximum of 5.5A at the motherboard slot, and we're showing a reading under 3A, which puts Inno3D's GeForce GTX 1060 Gaming OC in the green.

Of course, we have the data for measured currents as well:

The Cooler & Its Performance

The question remains: what kind of performance can we expect out of this budget card's cooler? According to the label on its backplate, we should be enjoying the benefit of HerculeZ Design!

Inno3D's backplate is screwed into the cooler from behind and into the PCB from the front. Spacers measuring almost 3mm tall allow some airflow between the board and backplate.

This isn't one of the beefiest coolers we've tested. It employs horizontally-oriented fins, two nickel-plated 6mm heat pipes above the base plate, and four nickel-plated 6mm pipes towards the card's rear. For cost reasons, the heat sink and base plate are both made from aluminum.

It deserves a positive mention that the heat sink for the MOSFETs is built into the cooler and does an excellent job. Unfortunately, the coils do not benefit from this implementation.

A measurement between 140 to 142°F (60 to 61°C) during our gaming loop, peaking at 145°F (63°C) in a closed case, shows that Inno3D's cooler works amazingly well.

Measurements taken with the backplate removed and the cooler mounted directly to the PCB show that Inno3D's thermal solution still has some performance in reserve. Results from our gaming loop and stress test reveal no abnormalities, except for extremely low temperatures around the VRMs thanks to an integrated sink responsible for drawing heat away from that area.

In both our gaming workload and stress test, Inno3D's thermal solution works the way it should, leaving us with one of the coolest boards in today's round-up.

Although the pressure to cut costs is apparent in several place, this cooler does a great job. Performance was clearly one area Inno3D refused to compromise.

Sound Results

A picture-perfect implementation of proper hysteresis is critically important to Inno3D. However, fan speed readings as high as 1400 RPM point to where this card's great cooling performance comes from. We think there's room to dial back a bit, since even at 1000 RPM there'd be enough cooling headroom to keep Nvidia's GP106 operating at under 167°F (75°C), while at the same time reducing noise quite a bit.

When the card is idle, its noise level is imperceptible due to a semi-passive mode. There's no point taking readings in that state.

Under load, a 35.7 dB(A) reading is neither loud nor really quiet, but instead average. If you're looking for a much quieter result, take advantage of the cooler's headroom and use a third-party tool like MSI Afterburner to dial in a lower fan speed.

A spectrum analysis turns up significant vibrations in the low-frequency range. Given this card's price tag, though, it's hard to expect perfection.

MORE: Nvidia GeForce GTX 1080 Roundup

MORE: Nvidia GeForce GTX 1070 Roundup

MORE: All Graphics Content

MSI GeForce GTX 1060 Gaming X 3G

The introduction of Nvidia's GeForce GTX 1060 3GB means we all have to pay more attention to naming. Not only does this model come equipped with half as much memory, but its GPU also sports 1152 active CUDA cores instead of the 6GB version's 1280. That means GP106-300 offers exactly 10 percent less theoretical compute power than the original GP106-400. Then there's the issue of on-board RAM. A measly 3GB could prove to limit current titles, even at 1920x1080.

These cards were first introduced in Asia, which doesn't come as a surprise since the Asian market has a different focus. Cards like this GeForce GTX 1060 "Light" will find plenty of satisfied customers among the millions of MMORPG and MOBA players.

Technical Specifications

MORE: Best Graphics Cards

MORE: Desktop GPU Performance Hierarchy Table

MORE: All Graphics Content

Exterior & Interfaces

The cover above the GeForce GTX 1060 Ti Gaming X 3G’s cooler is made of relatively thin plastic. It’s similar to the shrouds used on the larger cards in MSI's Nvidia-based line-up. A massive frame under the cover keeps the card stable.

At 1001 grams, this board isn't exactly light. Incidentally, it weighs in at just two grams less than the 6GB version. You can probably attribute the difference to the memory modules used on this model, which we'll discuss shortly.

The card’s other physical attributes mirror the 6GB version. Its dimensions are 28.4 x 13.5 x 3.5cm (L x H x D), spanning two expansion slots and standing relatively tall. The circumferences of the two dual-ball-bearing fans are also the same at 9.5cm.

The back of the board is hidden under a plate that consists of a single large piece. It has holes to allow air circulation, but doesn’t cool the card in any meaningful way since it’s not connected via thermal pads. The plate also adds 5mm to the back of the card, which can be a deal-breaker if you're butting up against a large cooler or memory modules.

Three nickel-plated heat pipes made from composite material are observable on the bottom of the GeForce GTX 1060 Ti Gaming X 3G. The two 6mm pipes travel all the way through, transporting waste heat to both parts of the cooler.

MSI's logo on the top of the card lights up. Again, the two 6mm heat pipes make a prominent appearance, as does an 8mm pipe. The eight-pin power connector sits on the end of the card, turned by 180°.

The rear of the card is completely open. The cooler’s fins are positioned horizontally, directing airflow towards the slot cover and the rear. This design has the advantage that a lot of its heated air gets blown out the back of your chassis.

The slot cover sports five connectors, four of which can be used concurrently in a multi-monitor setup. There’s one dual-link DVI-D connector without an analog signal, joined by two HDMI 2.0 connectors and three DisplayPort 1.4 outputs.

The rest of the slot cover is mostly taken up by vents. Airflow could have been improved by omitting the largely pointless DVI connector and including an adapter instead. This would have allowed for better cooling performance.

Board & Components

A quick look at the GeForce GTX 1060 Gaming X 3G’s board proves it to be pretty much identical to the 6GB version. All five phases are powered through the auxiliary connector. The memory modules have their own phase supplied through the motherboard’s PCIe slot.

The GPU power supply’s five phases are controlled by a uP9511P made by uPI Semiconductor Corp. Dual N-channel MOSFETs in the M381 line are used for the high and low side voltage converters. This setup eliminates the need for separate gate drivers, which saves both space and money.

The memory’s single phase has one small uP1641P also supplied by uPI Semiconductor Corp. It’s joined by a different voltage converter, the SM7320, which is also a dual N-channel MOSFET.

MSI uses the same coils we've seen for years. Their quality is just what you’d expect from a mass-market product in this category with decent encapsulation. That's to say they're better than Foxconn’s Magic coils.

An INA1221 is in charge of the card’s power, and we'll soon see that setting a high power target through the firmware doesn’t really benefit real-world performance. This is especially true if the power target is pushed as far as MSI takes it. Nvidia implements its own internal limiter that won't let overclockers get very far.

Only six of the eight available memory emplacements are populated with Samsung K4G41325FC-HC25 modules (32x 128 Mb). Their voltages range between 1.305 and 1.597V, depending on clock rate. Altogether, this amounts to just 3GB of graphics memory.

Two capacitors are added right below the GPU to smooth out voltage peaks. This design is similar to Nvidia's reference implementation.

Power Results

First, we'll look at GPU Boost clock rates. The factory setting initially yields an aggressive 1974 MHz, though that measurement drops slightly to 1949 MHz once the card is fully warmed up.

Of course, the voltage curve dips similarly, starting at 1.062V and ending at 1.043V, corresponding perfectly to the slightly lower GPU Boost step.

While we use several games with different rendering paths and quality settings, we're only testing at one resolution: 1920x1080. Benchmarking at higher resolutions just doesn't make sense on a card with 3GB of memory.

The GeForce GTX 1060 Gaming X 3G’s power consumption at idle and low loads turns out to be very similar to the GP106-400-based card. However, differences between the two boards increase as the load goes up. Doom doesn’t use the most power this time around; that honor goes to Metro: Last Light instead. Frame rates really suffer, though. The Witcher 3 is another example of a game that doesn't run smoothly unless the quality settings are significantly relaxed.

At playable settings, gaming power consumption is much more reasonable at around 102 to 103W. We take this as an indication that 3GB of graphics memory holds the card back more than we’d like. From here on in, we’re sticking with Doom to illustrate real-world performance.

The gray bar represents power consumption based on those load peaks that made it through our filters to the smoother curve. That bar doesn't have any practical significance since the peaks we measured are too brief for them to matter (even if the shortest-duration ones were already filtered out by this point).

Power Connector Load

Next, we drill down a bit more to look at how the loads are distributed between the two different 12V rails (motherboard slot and external power connector) during the realistic gaming load and stress test. Readings of 21W during gaming and 30W during the stress test show that the motherboard slot doesn’t see a lot of usage.

This is due to the fact that the memory and board components have the motherboard slot to themselves. The GPU draws all of its power through the auxiliary power connector. Now it's easier to explain why MSI uses an eight-pin connector, even though the GeForce GTX 1060 Gaming X 3G’s power consumption remains lower than 150W during all of our usage scenarios.

Here are the corresponding graphs for gaming and our stress test. Click on them for a larger version.

The PCI-SIG’s specifications only apply to current, meaning power consumption results on their own aren't comprehensive enough. Our readings put the motherboard slot just over 2.5A. Given a ceiling of 5.5A, this is most certainly on the safe side with lots of room to spare.

Of course, there are larger graphs for the current measurements as well.

The Cooler & Its Performance

The GeForce GTX 1060 Ti Gaming X 3G’s board is sandwiched between a heavy frame on top, and the aforementioned back plate on the bottom. Since these two components are screwed together, they basically create one massive and very solid part.

The frame is also screwed to the slot cover, hiding most of the board's top. Thermal pads connect it to not just the memory modules and MOSFETs, but also the two PWM controllers and one voltage converter responsible for the card’s other components.

The cooler is identical to what we found on MSI's GeForce GTX 1060 Gaming X 6G. It’s dominated by a nickel-plated heat sink that's screwed on in a way that presses its flattened fins onto the top of the cooler’s base. MSI doesn't use a massive plate with the heat pipes pressed into it for the GTX 1060, wheres the company did for its GeForce GTX 1070 and 1080.

Our new measurement and analysis software package enables us to examine benchmark results in new ways. The graph below shows the relationship between temperature and gaming performance. The actual frame rates during our strategically-chosen test run stay consistently at 79 to 80 FPS. The only aberrations are some small spikes due to the game.

This means that the GPU Boost clock rate isn't affected by a thermal limit. We know from our other benchmarks what really holds the GeForce GTX 1060 Gaming X 3G back. We’ll see the small valley that was produced by the game in our other graphs as well. This is how synchronized the measurements have become with our new software.

The thermal reading from below the GPU's package is consistent with the sensor output from within the GPU for our gaming loop and stress test. Meanwhile, the voltage converters’ and memory modules’ temperatures are well within the green range. Right underneath the memory modules, we measure temperatures of significantly less than 70 °C, even though those components are rated for up to 85 °C. The voltage converters come in at a cool 73 °C.

During the stress test, the area around the five voltage converters reaches 75 °C, and the hottest parts of the memory modules stay below 75 °C. MSI's cooling solution does its job, that's for sure.

Sound Results

The two large fans barely spin faster than their minimum RPM. Even during the stress test, they remain below 900 RPM. From 50 cm away, this isn't really audible.

The fans don’t even spin at idle, so noise results aren't necessary. During gaming, the GeForce GTX 1060 Gaming X 3G posts 29.3 dB(A). Its profile is dominated by deeper bearing noises, whereas the actual fan noise can’t really be heard above the card’s overall soft hum.

This result, combined with the temperature results, show that the card’s cooler is larger than it needs to be. Although the cooling performance is appreciated, everything comes at a price. The question whether that price is too high segues perfectly to our conclusions.

MORE: Nvidia GeForce GTX 1080 Roundup

MORE: Nvidia GeForce GTX 1070 Roundup

MORE: All Graphics Content

MSI GeForce GTX 1060 Gaming X 6G

After reviewing its GeForce GTX 1080 and 1070, MSI is presenting us with a third graphics card armed with the same cooler. Naturally, on a GeForce GTX 1060, it feels a little oversized. The big thermal solution's advantage is quite obvious though; it's probably going to be even quieter than either of the two higher-end models. Otherwise, this card looks pretty similar. The most apparent difference is its lack of an SLI connector.

Even if there's not a lot about the card's appearance to introduce, our performance results are still worth going into depth on. And there's an important caveat we need to present upfront:

We test all samples as they arrive, out of the box, exactly as most gamers use them. Bundled control software or third-party tools are deliberately avoided. Once again, MSI tried to exploit this policy by delivering a so-called golden sample with a modified firmware, just as it did in the case of our GeForce GTX 1080 Gaming X 8G. The sample employs higher clock rates (and perhaps a higher power target) than the BIOS you'd receive from an off-the-shelf board.Therefore, we cannot grant this card any award, no matter how good the technical implementation may be. In addition, we also have to disqualify this card from our comparative summary; it just wouldn't be fair to the other contenders.

Technical Specifications

MORE: Best Graphics Cards

MORE: Desktop GPU Performance Hierarchy Table

MORE: All Graphics Content

Exterior & Interfaces

Similar to its higher-end relatives, this card's fan shroud is made of relatively thin plastic. Underneath, you'll find a massive frame that improves the structure's stability.

Weighing 35.4 oz (1003 grams), the GeForce GTX 1060 Gaming X 6G isn't light (it's only about 100g lighter than the GTX 1080 version). The card measures 28.4cm long, 13.5cm tall, and 3.5cm wide, monopolizing two full expansion slots and plenty of area vertically, coming off the motherboard. This is due to the large fan blades, which stretch out to a diameter of 9.5cm.

The back of the board is covered by a single-piece plate with some ventilation holes. Unfortunately, there are no thermal pads connecting it to the circuit board, so the plate does nothing to help with cooling. Because of the backplate, though, you must plan for an additional one-fifth of an inch (5mm) of clearance to avoid collisions with your CPU's cooler, particularly in small form factor cases. A lack of SLI support means you don't have to worry about fitting two GeForce GTX 1060s back to back in a case.

MSI's logo on the top of the card lights up. Again, the two 6mm heat pipes make a prominent appearance, as does an 8mm pipe. The eight-pin power connector sits on the end of the card, turned by 180°.

The black and bright red colors follow MSI's corporate design, while, again, the card's shape resembles GeForce GTX 1070 and 1080 boards in the Gaming X line-up.

The rear of the card is completely open. The cooler’s fins are positioned horizontally, directing airflow towards the slot cover and the rear. This design has the advantage that a lot of its heated air gets blown out the back of your chassis.

The slot cover sports five connectors, four of which can be used concurrently in a multi-monitor setup. There’s one dual-link DVI-D connector without an analog signal, joined by two HDMI 2.0 connectors and three DisplayPort 1.4 outputs.

Where you don't find display outputs, the rest of the bracket features openings to facilitate airflow. If MSI chose to leave off the seldom-used DVI connector, there'd be even more space for ventilation. As you may have noticed, Nvidia went that route with its recent GeForce GTX 1080 Ti announcement, bundling a DP-to-DVI dongle for those who still need the interface.

Board & Components

All five phases are powered through the auxiliary connector. The memory modules have their own phase supplied through the motherboard’s PCIe slot.

The GPU power supply’s five phases are controlled by a uP9511P made by uPI Semiconductor Corp. Dual N-channel MOSFETs in the M381 line are used for the high and low side voltage converters. This setup eliminates the need for separate gate drivers, which saves both space and money.

The memory’s single phase has one small uP1641P also supplied by uPI Semiconductor Corp. It’s joined by a different voltage converter, the SM7320, which is also a dual N-channel MOSFET.

MSI uses the same coils we've seen for years. Their quality is just what you’d expect from a mass-market product in this category with decent encapsulation. That's to say they're better than Foxconn’s Magic coils.

Two capacitors are added right below the GPU to smooth out voltage peaks. This design is similar to Nvidia's reference implementation.

Only six of the eight available memory emplacements are populated with Samsung K4G41325FC-HC25 modules (32x 128 Mb). Their voltages range between 1.305 and 1.597V, depending on clock rate. Altogether, this amounts to just 3GB of graphics memory.

Power Results

Before addressing power consumption directly, let's take a look at the relationship between clock rate and voltage. In the case of MSI's GeForce GTX 1060 Gaming X 6G, this illustrates that frequency doesn't drop significantly during gaming, even after an imposed warm-up period.

Our observed GPU Boost rate falls from 2000 MHz to a still-respectable 1974 MHz over time and under load. The curve's slope is mirrored in our voltage readings. A 1.05V starting point pulls back ever-so-slightly to 1.043V by the end of our run.

The lowest GPU clock rate measured at idle is about 240 MHz.

We took our measurements using a variable low-pass filter, so we mention short load peaks only as a side note (see the grayed-out bar in the chart below), since those peaks are rarely relevant in practice.

Power Connector Load

Since MSI's GeForce GTX 1060 Gaming X 6G doesn't use the motherboard slot's 3.3V rail, we left it out of our chart.

The power distribution we observe shows a heavy bias toward the auxiliary power connector, which supplies just under 100W.

Here's the data for the gaming and stress tests:

Power consumption on its own only tells us part of the story. The PCI-SIG specifies a maximum of 5.5A at the motherboard slot, and we're showing a reading under 3A, which puts MSI's GeForce GTX 1060 Gaming X 6G in the green.

Of course, we have the data for measured currents as well:

The Cooler & Its Performance

Behind every successful high-performance graphics card is a well-designed cooler. In this case, MSI's implementation is pretty elaborate given the GP106 GPU's thermal ceiling. In contrast to MSI's GeForce GTX 1070 and 1080 Gaming X boards, the 1060 Gaming X relies on a closed frame that not only keeps the board nice and stable, but also helps cool the memory modules and VRMs.

Thermal pads provide the bridge necessary to transfer heat from the memory and power circuitry, though we would have preferred a thinner and more efficient design. After all, MSI uses low fan speeds that don't move much air.

The frame is screwed to the backplate, and ~3mm-tall spacers allow some air to flow between the plate and PCB.

As mentioned, this cooler is almost identical to the one used on the GTX 1070 and 1080 Gaming cards. It uses the same horizontally-oriented fins and heat pipe configuration. The direction they face is more or less irrelevant, since they're made of sintered composite material.

A massive nickel-plated sink takes the heat from those pipes and dissipates it out through aluminum fins.

The cooler does its job almost inaudibly. And yet it performs incredibly well, enabling readings around 145°F (63°C) during our gaming loop, and topping out at 149°F (65°C) in a closed case.

Our measurements with the backplate and stabilizing frame removed show that MSI won't let this card go up in flames. The gaming loop and stress test fail to turn up any apparent issues.

It becomes clear during our gaming test sequence that the lower five voltage regulator phases handle most of the load, while the memory's VRM remains fairly cool thanks to its less demanding task.

In the end, MSI's cooling solution does a great job (as it should, since it was designed for loads almost 100 percent higher).

Sound Results

Thanks to an exemplary implementation of hysteresis and fan speeds around 800 RPM, this card is almost inaudible.

When the GeForce GTX 1060 Gaming X 6G is idle, its noise is not measurable due to a semi-passive mode. Thus, we abstained from taking those readings.

Since this card's fans are the same ones used to cool MSI's GeForce GTX 1070/1080 Gaming boards (meaning they're almost overkill), our test chamber remains frighteningly quiet. An excellent 29.3 dB(A) result drops below the threshold of minimal background noise for a room inside a quiet apartment.

All we found were minimal vibrations in the low-frequency range caused by the slow-spinning fan motors. That's basically nit-picking though, since these vibrations aren't audible.

MORE: Nvidia GeForce GTX 1080 Roundup

MORE: Nvidia GeForce GTX 1070 Roundup

MORE: All Graphics Content

Palit GeForce GTX 1060 Super JetStream

Palit's GeForce GTX 1060 Super JetStream is virtually identical to Gainward's GeForce GTX 1060 Phoenix GS, with only a few small variations. One thing they share in common, unfortunately, is a lack of availability in the U.S.

Maybe this shouldn't come as a surprise. After all, the two cards come from the same factory. Where does this leave the Super JetStream in our tests, though, and could a few differences set Palit's card apart from Gainward's?

For a long time, the Super JetStream represented the top of Palit's factory-overclocking efforts. Only the higher-end GeForce GTX 1070 and 1080 have more aggressively-tuned GameRock Premium Edition models. In the GTX 1060 market, Palit decided against such a halo offering (and with good reason, as unnecessary fragmentation costs money).

Technical Specifications

MORE: Best Graphics Cards

MORE: Desktop GPU Performance Hierarchy Table

MORE: All Graphics Content

Exterior & Interfaces

The Super JetStream's fan shroud is made of black plastic with metal highlights. This cover is designed to be modular, as are the fans. Thus, a Palit card can easily be transformed into a Gainward model and vice versa by simply removing four screws.

Weighing in at 31 ounces (878 grams), the Super JetStream is moderately hefty, yet quite compact. Similar to the Gainward model, it measures only 24cm long, 11.7cm tall, and a substantial 4.7cm wide, practically requiring three expansion slots. Needless to say, you'd have a hard time wedging it into a mini-ITX-based system.

The back of the board is covered by a single-piece plate that lacks ventilation holes and does nothing to help with cooling. Because of the backplate, though, you must plan for an additional one-fifth of an inch (5mm) of clearance to avoid collisions with your CPU's cooler.

The top of the card features a backlit Palit logo with customizable RGB LEDs. A six-pin power connector is rotated 180° and positioned at the end of the PCB. This makes sense from a technical point of view, since the cooling fins don't need to be made any shorter.

A closer look at the heat pipes on the bottom reveals that they aren't embedded in a cooler block or heat spreader. Instead, they lay flat behind the base of the board (we already saw the same concept on Gainward's implementation). Whether this budget-oriented approach holds up any better in our tests remains to be seen.

Horizontally-oriented fins allow some waste heat to escape from an externally-facing slot bracket loaded with ventilation holes. This also applies to the card's other end, though, where hot air dumps out into the middle of your case. Good circulation is critical as a result.

The slot cover exposes five outputs, four of which can be used simultaneously to drive multi-monitor configurations. In addition to one dual-link DVI output (which lacks an analog signal), you also get one HDMI 2.0 interface and three DisplayPort 1.4-capable connectors.

Board & Components

The PCB that Palit and Gainward use features some prominent differences compared to Nvidia's reference design.

Right off the bat we see four power phases, just like Nvidia's card. Three are meant for the GPU and one corresponds to the memory. Two of the three GPU phases are fed by the auxiliary power connector. The third GPU phase and the one for the memory get power from the motherboard slot.

While, in theory, Palit's PCB offers the option to use one more power phase, the company went with a rarely-used and rather inexpensive NCP81174 for its PWM controller, which can only control up to four phases. Thus, the vacant emplacement you see in the image above could only have been populated by another parallel regulator circuit (via doubling).

The three GPU phases utilize a highly integrated Vishay SiC632, which combines the gate driver, high- and low-side MOSFET, and Schottky-diode in one convenient package. Meanwhile, the memory is driven by a pair of N-Channel MOSFETs, including one SM4503NHKP (30A, high-side) and one SM4377NSKP (50A, low-side), both manufactured by Sinopower. The MOSFETs are also fairly cost-effective choices.

Overall, the board reflects that Gainward didn't spend big on pricey components. Even the Foxconn Magic chokes are middle-of-the-road. They get their job done in a rather unremarkable way.

Two capacitors are installed right below the GPU to absorb and equalize voltage peaks.

Only six of the eight available memory emplacements are populated with Samsung K4G80325FB-HC25 modules. Each one adds 8Gb (32x256Mb) and operates at voltages between 1.305 and 1.597V, depending on clock rate. In total, they add up to this card's 6GB of GDDR5.

Power Results

The GPU Boost frequency and core voltage curves are particularly interesting to us. Our data shows that frequency drops significantly as the card heats up during a gaming session. Whether that's due to a higher-quality GPU or the clock rate's higher starting point compared to Gainward's version is a question we can't easily answer.

Otherwise, we observe a lot of similarities. The 2025 MHz we initially measured falls to an acceptable 1949 MHz under load. Voltage behaves similarly: we measured 0.975V on average, with readings varying between 1.025 and 0.9V. All of this points back to GPU Boost 3.0 and a relatively low power target that limits consumption to a maximum of 130W.

The lowest GPU clock rate measured with an idle card is about 253 MHz, just like Gainward's take on the same implementation.

We took our measurements using a variable low-pass filter, so we mention short load peaks only as a side note (see the grayed-out bar in the chart below), since those peaks are rarely relevant in practice.

One set of readings that does stand out is the noticeably higher gaming power consumption, especially at lower resolutions. Meanwhile, up top, Palit's GeForce GTX 1060 Super JetStream fares the same as Gainward's GeForce GTX 1060 6GB Phoenix "GS". Since Palit is slightly faster at 1080p, though, the results remain proportionate.

A low power target causes GPU Boost to cut the voltages dramatically during our stress test, resulting in power consumption numbers that are lower than what we measure during a worst-case gaming workload.

Power Connector Load

Load distribution from the four voltage regulators to their respective power phases isn't problematic at all. However, there's little headroom for overclocking, even if it wasn't already limited by the firmware's power target and voltage settings.

Here's the data for the gaming and stress tests:

Power consumption on its own only tells us part of the story. The PCI-SIG specifies a maximum of 5.5A at the motherboard slot, and we're showing a reading of about 5A, which puts Palit's GeForce GTX 1060 Super JetStream in the green (though a bit higher than Gainward's card).

Of course, we have the data for measured currents as well:

The Cooler & Its Performance

Although we already covered Gainward's cooler, it's worth discussing Palit's as well since there are a couple of details that differ. Like the backplate Gainward uses, this one is purely aesthetic; it doesn't help cool the PCB at all. Four screws secure it to the top of the board, shielding the card from curious customers.

The fan shrouds that Gainward and Palit use look a little different. Gainward's card is equipped with slightly more powerful fans (6.6W, 2700 RPM max.), while Palit's Super JetStream only includes 4.8W models, which top out at 2500 RPM. In addition, the 9cm blades employ a different geometry, which in Palit's case is slightly wavy. Their surfaces is smooth, without grooves.

It's also worth mentioning that the fans rotate in opposite directions. This should help improve efficiency due to more chaotic air flow.

Again, the delicate-looking cooler sports horizontally-oriented fins, as well as a quartet of 4mm heat pipes made from sintered composite material. Those pipes are not integrated into the bottom of the cooler. Instead, they merely touch the back of the thin base plate. Expensive copper is nowhere to be found.

We're glad to see that the base plate has a specific area to help cool the MOSFETs. However, the coils have to live without the benefit of active cooling.

Registering between 162 to 163°F (72 to 73°C) during our gaming loop, and a maximum of 169°F (76°C) in a closed case, the cooler does a fair job. It would almost assuredly perform better if the heat pipes were attached more securely, though. The stress test shows a similar picture.

The temperatures we measure on Palit's GeForce GTX 1060 Super JetStream are one to two Kelvin above Gainward's, which could be due to the higher power numbers we observed and its weaker fans.

Measurements collected with the backplate removed also reveal acceptable values. A peak of approximately 180°F (82°C) from the VRMs is reasonable, in our opinion. Interestingly, the different fan geometry clearly helps cool the PCB more efficiently.

With readings around 189°F (87°C), our stress test doesn't reveal any serious problems.

Palit's cooling performance is basically acceptable. We get the impression that one member of the finance team was assigned to each engineer, keeping costs in check more than enthusiasts like to see.

Sound Results

Hysteresis is not an issue for most manufacturers. However, Palit implemented a fan curve with too-low of an initial rotational speed for the transition between passive and active mode. As a result, the fans simply stop, and need to be restarted repeatedly. Even the best hysteresis won't help when a manufacturer overestimates the quality of its components this grossly.

A minimum value of 400 to 450 RPM (~600 RPM as a start-up pulse) would have been necessary, corresponding to about 18 percent as a PWM controller signal. Knowing this helps us manually modify the fan curve accordingly, thus getting rid of the annoying start-stop cycles.

When the card is idle, noise is not measurable due to a semi-passive mode. A reading of 35.8 dB(A) under load is acceptable, but not great. So much air has to be moved through the thermal solution because Palit's cooler isn't as efficient as it should be. Spending a few dollars more would have gone a long way. On the bright side, we measure very little low-frequency bearing noise. What remains is the sort of acoustic energy that's easily absorbed by a case with sound insulation.

MORE: Nvidia GeForce GTX 1080 Roundup

MORE: Nvidia GeForce GTX 1070 Roundup

MORE: All Graphics Content

Introduction & Overview

What would happen if your SSD lost its memory? Contrary to popular belief, DRAMless SSDs are not new. The first consumer SSDs shipped without a DRAM cache and experienced a stuttering issue, so the industry addressed slow performance with DRAM. Later, SandForce released controllers that leveraged a small amount of memory tucked inside the flash processor, a technique that we still find in most SSD controllers. We're examining several new retail DRAMless products, along with two next-generation models featuring NAND we've never tested before. By the end of this article, we hope to know if modern DRAMless SSDs take us back to square one, or if these products extend from where SandForce left off.

For various reasons, over the years several companies have invested in technology to eliminate the DRAM cache. Some of the companies don't manufacture volatile memory--only Samsung, Sk hynix, and Micron (Crucial) manufacture both NAND and DRAM. The other SSD manufacturers have to purchase the memory from a competitor that is also trying to sell SSDs. The DRAM also consumes more power than flash, is subject to wild price swings, sometimes is in short supply, and is a critical component that most companies do not control.

Your computer keeps frequently used and application data in memory, which you may know as SDRAM, DDR, DDR3 or DDR4. SSDs use memory a little differently, though it can be just as important. SSDs do not store data the same way as an HDD, primarily because NAND has a limited lifespan. Your operating system was built on the assumption that it would store data on a hard disk drive, so it steers data to specific locations (addresses) for storage.

The SSD controller shuffles the data around to different addresses to ensure the flash wears out evenly, but the operating system isn't aware of the data movement. The SSD also conducts garbage collection operations to provide enough free blocks to ensure high performance, along with several other background tasks that result in a completely different set of internal data addresses.

In effect, the operating system has one map of data addresses where it thinks the data is located, which the SSD considers the logical map, while the SSD has its own physical map that reflects where data is actually located. The SSD maintains the two separate maps, and it references or updates the maps every time the operating system needs to access data. The Flash Translation Layer (FTL) coordinates all of the processes.

SSDs will keep all or a portion of the map in DDR2 or DDR3 (usually). DRAM is much faster than NAND, so the SSD can access the map quickly to increase performance. DDR-type memory loses data when there isn't any power, but NAND stores it even when the power goes out. The SSD keeps a copy of the map on the NAND to reduce the chance that it will lose the map from a power loss. (Technically, the SSD doesn’t lose the data, it just loses the map, so it can't find it).

There are a few different approaches to eliminating DRAM. We only know of a few because the deep inner workings and algorithms tend to be closely held trade secrets. A common method is to build a small amount of memory into the controller. The Phison S11 controller we have in our tests pool features 32MB of SRAM built into the controller, but that is a very small amount of memory compared to an external module. Other techniques include compressing the flash translation layer map (essentially, it is a complicated spreadsheet, so it compresses easily) or caching a portion of the map in system memory (HMB). The SSD controller accesses as little as 8 percent of the map frequently, so there are ways to reduce the performance loss.

Unfortunately, DRAMless SSDs also have a sinister side. Updating the map directly on the flash requires small random writes, which takes a bite out of the SSD's endurance. This is a particularly vexing issue with low endurance planar 2D TLC NAND flash. At Computex last June, one SSD vendor told us about an OEM 2D TLC SSD that will burn through the rated endurance in a little over a year. The SSD has to last a year because of the notebook's one-year warranty, but anything beyond a year's worth of use is up to the user to fix. Tactics like that are the driving forces behind putting cheap DRAMless SSDs in $500 notebooks.

Technical Specifications

We have eleven SSDs on the grid. Other than our two prototype reference design parts, the other nine are sold online and in retail stores. The retail parts come from four different series and three different manufacturers. We will go over each model individually. Let's start with the cutting-edge development products armed with 3D NAND before we move on to the details of the retail products.

MORE: Best SSDs

MORE: How We Test HDDs And SSDs

MORE: All SSD Content

128GB Benchmark Results

Product Comparisons

We only include DRAMless models in our performance breakouts. This is a special class of products that falls outside of traditional SSDs, and in the future, we will see more products in the DRAMless category. I would certainly hope not, but the push to create lower cost SSDs brings about some interesting products. Just like the transition to NVMe, DRAMless SSDs will grow and eventually most entry-level SATA products will ship without DRAM.

We have three separate controllers in the 120GB-class, though two of the SSDs are paired with roughly the same NAND. Toshiba (OCZ) and SanDisk produce NAND in the joint Flash Forward venture. Each company takes possession of the finished wafers, but there are some subtle processing differences. The companies may cut the die differently and use a different process to bin or sort by quality. Controller programming is different for each of the three products. The OCZ TL100 uses an in-house Toshiba controller. SanDisk tapped SMI for the DRAMless drives, but the SSD Plus 120GB uses the older SM2246XT controller with 15nm MLC. The Z410 uses the new SM2256XT controller but pairs it with 15nm TLC.

Sequential Read Performance

To read about our storage tests in-depth, please check out How We Test HDDs And SSDs. We cover four-corner testing on page six of our How We Test guide.

It shouldn't shock anyone that the SanDisk SSD Plus outperforms the two TLC-based products in many of the tests. The SLC buffer in the TLC products will outperform the MLC-based SSD Plus in some tests, but it only operates at full speed for a few seconds. The sequential read test shows that the two-package MLC drive still delivers higher performance when you request data back from the drive.

Sequential Write Performance

The MLC flash in the SanDisk SSD Plus 120GB doesn't lose performance as fast as the TLC drives. Testing DRAMless SSDs is difficult due to the rapid write performance decline and how low the performance drops. We stopped reviewing 128GB-class products over a year ago. 256GB-class products often cost only a few more dollars, so 128GB-class SSDs aren't a good value.

The wave in the test results is due to the SLC buffer filling and emptying. We can't control at which point in the test the buffer ebbs and flows. High-capacity SSDs are more predictable because the buffer is larger, but the low capacity SSDs have a much smaller buffer so the user experience is less enjoyable.

Sustained Sequential Write

We were surprised to see the OCZ TL100's high performance in the HD Tune Pro sequential write test. Performance should have dropped, but it didn't on the first pass. The SSD Plus didn't drop either, but its MLC flash could only muster a 210 MB/s sustained write speed. The Z410 is the biggest loser in this test. The native TLC performance was just under 100 MB/s, but the part we really worry about is the massive 43ms latency spike during this very simple test. Even though it was a single outlier, during normal use, you would think the system locked up and was about to kick out a blue screen.

Random Read Performance

With three different controllers in the mix, there are numerous angles to tackle in this section. Toshiba has the upper hand here because it manufactures both the flash and the controller, so it can tune both to work together more effectively. SanDisk used a third party controller, so some of the more advanced and targeted customization isn't possible. Toshiba managed to deliver better random read performance at QD1, but the two SanDisk DRAMless SSDs strike back at QD2. 

Random Write Performance

The random write performance tests came out exactly how we expected. The SSD Plus 120GB with MLC NAND walked away from the two TLC-based drives, but all of the results are equally unimpressive.

80 Percent Mixed Sequential Workload

We describe our mixed workload testing in detail here and describe our steady state tests here.

DRAMless SSDs deliver similar sequential mixed performance compared to TLC-based products. In this test, there really isn't that much difference compared to a DRAM-equipped product, like the Adata Ultimate SU800 we recently reviewed.

80 Percent Mixed Random Workload

Mixed random performance is much lower than a TLC-based drive with a DRAM cache. The benefits of increased random performance are difficult to quantify in words. When you transfer files at home, Windows delivers a pop up that says how fast the file transfers. This makes increased sequential performance easy to spot. Random performance is less obvious, but you can feel when a high-performance SSD is in a system.

Sequential Steady-State

The SSD vendors didn’t design the three products in this chart for professional applications or heavy write environments. The test shows how low your sequential write performance can drop with a drive nearly full or during a heavy session of data writes.

Random Steady-State

The SanDisk Z410 is the only model with consistent performance. It delivers around 300 random read IOPS in steady-state (a measurement of worst-case performance). DRAMless SSDs are synonymous with "hard disk replacement," which is the term the SSD industry prefers. The three products deliver higher performance than an HDD, but you can't buy a 120GB hard drive, either.

PCMark 8 Real-World Software Performance

For details on our real-world software performance testing, please click here.

The three small capacity DRAMless SSDs take shots at the top of the performance list. The different controller and flash architectures lend a helping hand in some tests, but the configurations aren’t ideal for all consumer workloads. 

Application Storage Bandwidth

The SanDisk SSD Plus with the SMI SM2246XT controller and Toshiba MLC flash outperforms the OCZ TL100 and SanDisk Z410 with TLC flash, but only by a small margin.

PCMark 8 Advanced Workload Performance

To learn how we test advanced workload performance, please click here.

Coming out of a heavy sustained workload, the MLC-powered SSD Plus recovers faster than the other products. The other two products recover, but at a much slower pace, and they fail to reach peak performance again during the trace-based test.

Total Service Time

We expected the OCZ TL100 120GB to outperform the SanDisk Z410 120GB in this test, but that didn't happen. The SSD Plus 120GB with MLC flash completed the application tests much faster than the other products. To be very clear, the SSD Plus didn't deliver stellar performance. The SSD Plus wouldn't look as good in comparison with just about any modern DRAM-equipped SSD.

Notebook Battery Life

There is a very wide gap between the Z410 120GB, which has the longest battery life, and the OCZ TL100. The gap is roughly half an hour. It's difficult to imagine the large difference with so few components. 
MORE: Best PC Builds

MORE: How To Build A PC

256GB Benchmark Results

Comparison Products

The 256GB-class is the most populated. Here we see both current and next generation reference designs that predict the future of the DRAMless category. 3D NAND flash will breath new life into DRAMless products, but low yields and a slow 3D production ramp means 2D planar products will dominate the first quarter of 2017 for every NAND manufacturer other than Samsung.

There are two drives with MLC in this capacity class. The Phison S11 reference design uses Micron 3D MLC. This is the first time we've tested this flash. The SanDisk SSD Plus also features MLC, and we have a Phison S11 with second-generation 48-layer Toshiba BiCS TLC. BiCS is the trade name for Toshiba and SanDisk's 3D flash. The joint venture has announced 64-layer BiCS 3 and a new manufacturing facility. Apple uses 48-layer BiCS in the new iPhone, so it is in production, but this is the first time we've tested it.

The OCZ TL100 240GB, the largest capacity in the TL100 series, and the SanDisk SSD Plus 240GB and Z410 240GB populate our test pool. 

Sequential Read Performance

To read about our storage tests in-depth, please check out How We Test HDDs And SSDs. We cover four-corner testing on page six of our How We Test guide.

Performance increases faster than the cost in this capacity class. If you just need a boot drive for a system without a large game or multimedia collection, the 256GB-class is one of the best values. That carries over to the DRAMless solutions, as well.

Even though we have two next-generation products in the chart via the two Phison S11 reference designs, the Toshiba OCZ TL100 240GB performs a little better during the sequential read test. It's important to remember the two Phison reference design SSDs are not completely optimized. It takes time to fine-tune products. The reference designs are essentially a guide but are not representative of the final products.

Sequential Write Performance

The increased parallel writes to the larger NAND packages helps the TLC-based products deliver more write performance compared to the smaller capacity SSDs on the previous page. However, this capacity class is still small enough for background activity to skew the results. The SSD hits a heavy steady-state-like performance during these workloads due to native TLC performance.

Sustained Sequential Performance

The OCZ TL100 and Phison S11 with next generation Toshiba BiCS NAND didn't show any performance drop off even though they both use TLC. We ran the test several times in full span accurate mode, which is a fancy way of saying we wrote to the entire addressable space of the SSD. The performance is similar to what we observe on Samsung TLC-based products, which is good news for users. This test uses a single queue depth with 128KB blocks.

The Phison S11 controller paired with Micron 3D MLC doesn't drop, but it shouldn't because it uses MLC NAND. The two SanDisk SSDs drop even though the SSD Plus specifications (found on Newegg) state the drive uses MLC flash. Given the variable BOM, this drive may ship with TLC or MLC, and like others who purchase this drive, we have no idea what the SSD Plus actually ships with. Both of our SanDisk 240GB DRAMless SSDs use the same SMI SM2256XT controller with LDPC error correction. SMI designed the controller specifically for use with TLC NAND flash.

Random Read Performance

Random read performance is one of the most important metrics for modern software. Compared to the 120GB products, the 240GB-class DRAMless products all see an increase in this test at low queue depths. This is the saving grace for users running a light workload, and it makes these SSDs a viable solution. 

Random Write Performance

We've been unimpressed with Micron's 3D TLC since its launch, but it looks like the 3D MLC isn't too bad. I'd still choose Samsung V-NAND TLC over Micron's 3D MLC, but the lack of comparable samples with a DRAM cache skews my opinion.

All of the TLC and (maybe) TLC products bomb the random 4KB write test. We're certain now that the SanDisk SSD Plus 240GB uses TLC flash. The sample we used for this test is lacking, but it does deliver more random write IOPS at QD1 than a hard disk drive.

80 Percent Sequential Mixed Workload

We describe our mixed workload testing in detail here and describe our steady state tests here.

Again, we are surprised to see some of the products perform so efficiently in the sequential mixed workload test. The performance is similar to mid-tier SATA products sporting a DRAM cache. The two SanDisk drives fall outside of this realm and deliver poor mixed sequential performance. The two drives share very similar hardware, but the firmware is different, which is why we see two different results from such similar products. 

80 Percent Random Mixed Workload

The two SanDisk DRAMless SSDs fall behind the other SSDs during this test. The Micron 3D MLC NAND lags behind the Toshiba 3D NAND and OCZ TL100. This most likely goes back to pre-production versus production firmware. The retail TL100 240GB delivers the highest performance at low queue depths, and it appears to be one of the better DRAMless SSDs.

Sequential Steady-State

None of the DRAMless products are built for strenuous workloads, so the two steady-state tests don't provide us with a lot of useful information. We do get a look at worst-case performance if your drive is nearly full, which we wouldn't recommend with any of these products.

Random Steady-State

The 240GB DRAMless products don't provide a significant random 4KB steady-state increase over the 120GB models.

PCMark 8 Real-World Software Performance

For details on our real-world software performance testing, please click here.

Only one product really stands out from the rest. The Phison S11 with Toshiba 3D BiCS2 TLC consistently delivers enough performance to propel it to the top, or second from the top, spot in the chart. Phison works closely with Toshiba, so we could assume the S11 has undergone more optimization. It will be very interesting to see how high-performance BiCS SSDs compare to Samsung products if Toshiba and SanDisk bring any of those products to market.

Application Storage Bandwidth

The three retail products deliver nearly identical performance under realistic workloads using real-world software applications.

PCMark 8 Advanced Workload Performance

To learn how we test advanced workload performance, please click here.

Even after a cooling off period (idle time), the 256GB DRAMless products all deliver erratic performance. Ideally, the lines would all appear smooth on the chart, which is a sign of consistent performance.

Total Service Time

The Micron 3D MLC suffers from high latency after heavy use. We observed the same condition with the Crucial MX300 and Adata Ultimate SU800 using Micron's 3D TLC. This has been a real issue for Micron that the company has had a difficult time taming. It's now clear that high latency has become a consistent theme regardless of which controller the manufacturer uses with Micron 3D.

The chart shows the other products perform much better, but compared to many mainstream (and even entry-level) DRAM SSDs on the market, the results are very poor.

Notebook Battery Life

When we first received the MX300 750GB Special Edition with Micron 3D TLC, the drive wouldn't complete our MobileMark 2014.5 test due to high latency. Crucial held the reviews to address the high latency issue. Two weeks later, the company released an updated firmware that allowed the drive to pass the test. The Phison S11 doesn't appear to be as optimized for the new Micron 3D flash as it is for the Toshiba BiCS 2 3D TLC.

Somehow, the SanDisk SSD Plus provided more battery power time than the other retail products. It delivers nearly a half an hour of extra battery life over the SanDisk Z410. The difference comes down to the firmware. The OCZ TL100 delivered poor performance during this test once again.

MORE: Best Cases

MORE: All Case Content

512GB Benchmark Results

Comparison Products

The 512GB-class products are all retail units available for purchase. The SanDisk SSD Plus and Z410 appear again, this time with 480GB of capacity. The Patriot Spark joins the SanDisk duo and is the only product that makes the full LBA range available to the user. Running the Spark with more than 480GB of data could pose a serious performance issue. Most SSDs will use any available space as spare area, but this drive doesn't have a reserve for background activities if the drive is nearly full.

The SSD Plus is the only product in this review with a DRAM cache. We already had the drive tested and the charts compiled before we realized this DRAMless SSD actually has DRAM (at this capacity). We left it in as a comparison point. 

Sequential Read Performance

To read about our storage tests in-depth, please check out How We Test HDDs And SSDs. We cover four-corner testing on page six of our How We Test guide.

This version of the SanDisk SSD Plus uses a Marvell Dean 4-channel flash processor. This is the same part used in the Crucial MX300, SanDisk X400, and a few other new mainstream SSDs. The Spark joins the SSD Plus with acceptable sequential read performance, but the Z410 trails both at low queue depths.

Sequential Write Performance

The Patriot Spark 512GB has a rough time in our sequential write test. In the next section, the Spark gives us a misleading result during a write to the full usable space of the drive. In this test, following other disk activity, the Spark trails the other DRAMless SSDs.

Sustained Sequential Write Performance

This is where things get interesting. The HD Tune Pro write workload is the second test we perform when new SSDs hit the lab (the first is an HD Tune Pro read workload across the entire LBA range). The Spark 512GB didn't drop to native TLC speeds on the first pass, but it provided 150 MB/s during the second pass.

The two SanDisk SSDs deliver different levels of performance. The SSD Plus has a DRAM cache and a newer generation controller, so we can't chalk all of the performance variances up to the cache alone.

Random Read Performance

A DRAM cache increases random performance more than sequential performance, so it didn't come as a surprise when the SSD Plus 480GB outperformed the other two products. The Patriot Spark delivers better random performance at high queue depths compared to the Z410, but both DRAMless SSDs start out with nearly equal QD1 performance.

Random Write Performance

The SanDisk Z410 nearly drops off the chart. The Z410 appears to enter into a panic mode, which is never a good sign for flash-based storage. The result still weighs in at nearly 600 IOPS, but the SSD paused at some point during the test. The two other products perform as expected.

80 Percent Sequential Mixed Workload

We describe our mixed workload testing in detail here and describe our steady state tests here.

The increased parallel operations, which is the benefit of more NAND on the high capacity models, allows the three 512GB-class SSDs to deliver a significant performance increase over the models with less capacity. This is really the starting point for most of our readers as far as capacity and pricing go. It’s nice to see that these drives still deliver more performance than a hard disk drive (which has DRAM) even with the DRAMless disability in the architecture.

80 Percent Mixed Random Workload

That doesn't mean you should feel inspired to buy one of these products. All three would make a poor boot drive. The SSD Plus 480GB is the only model that we would consider a true entry-level SSD. You could use the others as low-cost mass storage with comparable sequential write performance to modern 7,200 RPM 2.5" disk drives.

Sequential Steady-State

It's interesting to see the different programming involved with the Z410 in different capacities. Normally we would expect the larger drive to deliver better performance across the board, but DRAMless products manage to turn the additional capacity into a disadvantage. It's important to understand what the cache does; it stores the LBA map for faster access. The size of the map swells as the amount of data on the drive increases. This drive has to manage a map that is four times larger than a 120GB model and twice the size of the 240GB model. That makes any steady-state level workloads frustratingly impossible.  

Random Steady-State

There is a wide gap between the three products during the steady-state random write test. The DRAM in the SSD Plus really helps here. The DRAMless Z410 shows the small amount of cache embedded in the controller caching the map, but it delivers inconsistent performance. The Spark has consistent performance but delivers it sparingly.

PCMark 8 Real-World Software Performance

For details on our real-world software performance testing, please click here.

The SSD Plus uses nearly the same configuration as the X400 we tested a few months ago. It doesn't perform as well with applications and time-based tasks, but it serves as a good way to compare an entry-level SSD's performance to the two DRAMless retail products. In some of the tests, we see the difference in the split seconds, but in many, we see the difference in seconds.

It's important to remember that under the RGB lights and fancy components a PC is just a tool. You always want your tool to operate as efficiently as possible. You don't change the job if it takes several seconds longer than it should, you change the tool.

Application Storage Bandwidth

Seconds add up throughout the day. Productivity is important, even if it comes in the form of a few hundred extra megabytes a second.

PCMark 8 Advanced Workload Performance

To learn how we test advanced workload performance, please click here.