Sign in with
Sign up | Sign in
AMD's 4x4 Platform & Athlon 64 FX-70 - Brute Force Quad Cores
By ,
1. Athlon 64 FX-70 Family Kicks Off Double-Socket Dual Cores

The Core 2 Duo processor launch on July 14 marks a turning point in the CPU world. It was on this day that AMD lost its dominance over Intel in the desktop processor space, a lead that had lasted for almost three years. Simply put, Core 2 Duo processors are faster and more energy-efficient than the Athlon 64 product family. While energy efficiency has to be evaluated for complete systems rather than only by processors, AMD still offers lower entry-level pricing. Clearly, AMD had to go back to the drawing table to come up with an adequate answer to Intel's challenge. Ladies and gentlemen, here it is: the 4x4 Quad FX platform, with dual-socket Direct Connect architecture for enthusiasts.

It is no secret that Intel has had a head start of at least nine months when it comes to manufacturing technology. All Core 2 Duo chips, the latest Xeons, and many of the Pentium D processors are based on the top-notch 65 nm process. Smaller transistor structures allow for lower energy consumption, more features (such as huge cache memories) and more processing cores on a chip. It is for this very reason that there is an Intel quad core (using two Core 2 pieces), but no AMD counterpart. The fact that AMD's current 90 nm silicon-on-insulator process is more advanced than the Intel equivalent doesn't help - a 90 nm quad core would simply be a huge (and expensive) piece of silicon.

AMD will introduce its first generation 65 nm products on December 5, which paves the way for a future quad core CPU. The initial 65 nm processors will target the entry level and the mainstream, though, and we will have to wait several more months before high-end 65 nm AMD processors hit the shelves. The 65 nm quad core Agena FX, with dedicated L2 caches and a unified L2 cache, is not expected before summer 2007. Until then, AMD had to come up with a solution to keep Intel's pace and show off innovation, because the first 45 nm Intel processors are very likely to arrive as soon as late 2007.

Quad FX is AMD's solution to the quad core issue, but what is it all about? AMD decided to appeal to the uber-enthusiast by offering a dual-processor platform that addresses the needs of so-called "megatasking environments". In short, that means doing everything you want at once. The platform technology could conveniently be taken from the Opteron world: Quad FX consists of a dual Socket 1207 motherboard using the beefed-up Nvidia nForce 680a chipset, and a pair of Athlon 64 FX-70 processors. Unlike the Opteron world, the Socket 1207 Athlon 64 FX processors do not require registered memory; they run on conventional DDR2-800 DIMMs - two per processor. To make the whole thing really attractive, AMD offers FX-70 bundles, which always consist of two FX-70, FX-72 or FX-74 processors (corresponding to speeds from 2.6 to 3.0 GHz).

The real question, of course, is if the result is worth the effort. We've put one of the first Athlon 64 Quad FX-74 systems to the test - and found out that this question is not that easy to answer.

Join our discussion on this topic

2. Who Needs Two Processors And Four Cores Anyway?

AMD targets the top of the line enthusiast and the real power user. No, not just "real", AMD means "real real". The buzzword for this mega platform concept is "megatasking", which AMD defines as follows:

  • Running multiple, heavily multi-threaded, processor-intensive applications simultaneously
  • Running digital media applications
  • Playing future multi-core games

The AMD Advantage

Most of the tasks described above will run properly on a powerful dual core processor today, but not necessarily in the future. While the general rule for multi-threaded environments is "the more cores the better", an increasing core count - four and more cores per processor - will create tremendous demand for processor interface bandwidth. Large amounts of data will need to be transferred to and fetched from the main memory as quickly as possible.

Intel will upgrade its 266 MHz quad-pumped FSB1066 processor Front Side Bus to 333 MHz (FSB1333) early next year with its Bearlake chipset (X38 for enthusiasts and P35 for upper mainstream). This which will provide a faster data path for dual and quad core CPUs to communicate with each other and with the main memory and motherboard core logic.

However, AMD has always had an advantage when it comes to memory efficiency, because each Athlon 64 processor has its own integrated, high-speed, low-latency memory controller. The first Athlon 64 processors supported DDR400 (Socket 754) or dual channel DDR400 memory (Socket 939). The current generation of Athlon 64 and Athlon 64 X2 processors on Socket AM2 runs dual channel DDR2-800 memory. A CPU core can fetch requested data directly, without the detour through a memory controller that is part of the motherboard core logic.

Why is this important? Well, it clearly doesn't seem to be today at least. This is mostly because Intel has the 65 nm manufacturing process advantage, which allows the company to fit larger L2 caches into the processor. Large and efficient caches allow for adequate performance, because they're capable of compensating for a memory subsystem that might not be state of the art. Efficient and shared caches (used by two or more cores at the same time) also help to reduce memory access.

As soon as Microsoft Vista arrives, things will change a bit. Windows XP is unable to tell physical processors apart, and it will distribute the processing of threads to all CPU cores as they are available. Whether the cores are located on one, two or multiple CPUs is irrelevant here, which can cause inter-processor thread-switching. This should be avoided, as it might entail the relocation of thread data as well - isn't that a nice new bottleneck!

Threads that were processed on one physical core might be handled by a different unit the next time they are being executed. If you now think of an Intel quad core created by putting two dual core processors into one physical processor package, you will realize that the Front Side Bus is the only way for inter-processor communication and main memory access.

In a worst case Windows XP scenario, processing unit A has to wait until unit B completes memory access. Then A accesses the memory to fetch data, which it stores in its L2 cache to provide it locally for processing. If, however, Windows assigns the thread to CPU B, it will have to fetch the current data from A's L2 cache, causing additional Front Side Bus traffic. For coherency and performance reasons, data cannot be fetched from the main memory again at this point, since it was already processed. In the end, all of the elements involved are slowed down by this maneuver.

Windows Vista Ultimate Edition will be able to tell processors or nodes apart from simple processing cores. This allows the operating system to assign threads in a more resource-efficient manner: one large task can be executed exclusively on CPU A, while another huge workload runs autonomously on CPU B. Inter-processor task switching is eliminated due to the enhanced hardware awareness of Vista Ultimate, and performance will scale much better with increased core count per processor.

Now, let's go back to the beginning: why is this so important? Remember that an Athlon 64 processor has its own memory controller, and thus a higher memory bandwidth per processing core. As soon as intelligence is added to the multi-threading/multi-tasking game, AMD's Quad FX platform will be able to show some serious muscle that it cannot in today's operating environments. Again, though, you have to bear in mind that we're talking about serious workloads that exceed what you and I typically do with our PCs.

3. Buying FX-70 Means You Buy Two Processors

The overall concept behind Quad FX is clear: "if we cannot provide a real quad core now, let's go and build something bigger with what we have." The result is three new processor products, all based on existing technology. If you purchase an Athlon 64 FX-70 ($599), FX-72 ($799) or FX-74 ($999), you will be buying two physical processors running at 2.6 GHz, 2.8 GHz or 3.0 GHz, respectively. Socket AM2 does not support more than a single processor socket, so AMD simply pulled the professional Socket 1207 off of the Opteron shelf to power Quad FX. So far, there is only one motherboard that will be available to support the new FX-70 series, but before we get to it, let's dig into Quad FX step by step.

Still The Good Old 90 nm Windsor/Egypt Dual Core

There is not much news on the CPU front: AMD uses existing designs and adjusted packages and ratings for Quad FX. Windsor is the codename for the Athlon 64 X2 and Athlon 64 FX dual core at 90 nm with 1 MB L2 cache per processor core. Egypt is the Opteron counterpart, which offers two more HyperTransport links and requires registered DDR2 memory. Apart from that, these two parts are like close siblings.

AMD deploys dual stress liner technology (DSL) in its silicon on insulator (SOI) process. DSL allows the chips to run higher clock speeds, and the insulator layers are used to oppose electromigration, which causes leakage currents and increases processor temperatures. Using DSL SOI, AMD was able to hit the 3 GHz mark with its latest products. However, there is not much headroom left: even overclocked to just slightly over 3.1 GHz, our test system crashed.

AMD FX-70 Models: Available At Up To 3 GHz

Some people didn't expect AMD to release the Athlon 64 FX-62 at 2.8 GHz until spring, but now AMD has proven that the 90 nm process is good enough to reach even higher clock speeds. The FX-74 is specified to run a core clock speed of 3.0 GHz and thus is the fastest AMD processor product ever.

Model Clock HT Speed Package Voltage Max. Case Temp TDP L2 Cache
FX-70 2.6 GHz 1 GHz bi-directional LGA 1207 1.35-1.4 V 55-63°C 125 W 1+1 MB
FX-72 2.8 GHz 1 GHz bi-directional LGA 1207 1.35-1.4 V 55-63°C 125 W 1+1 MB
FX-74 3.0 GHz 1 GHz bi-directional LGA 1207 1.35-1.4 V 56°C 125 W 1+1 MB

4. Two Processors = Double The Power Requirements

More processors is not necessarily a blessing once you start taking energy requirements and cooling into account. A single Athlon 64 FX-62 may run acceptably hot, but two of these power plants, and all of their required voltage regulators, will definitely heat up your computer much more. It is not surprising that the Quad FX machine set new records in terms of power consumption: we measured an idle power of 426 W for the FX-74, and 417 W for the FX-72. Both were running in our test system with a 1,000 W power supply, dual GeForce 7900 GTX graphics, and two hard drives.

An idle power of 261 W sounds much better, which is why we strongly recommend turning on AMD's Cool & Quiet feature. When enabled, during periods when processors run under low load or are idle, Windows XP (and Vista) will throttle the CPU core clock speed to 1.0 GHz and the system will reduce the CPU voltage from 1.4 V to 1.175 V to save power and reduce cooling noise. To use Cool & Quiet, you will need to install AMD's latest Windows XP processor driver, enable Cool & Quiet in the BIOS, and switch Windows XP to portable mode.

Step Down: 1 GHz Clock In Cool & Quiet Mode

The 3.0 GHz Athlon 64 FX-74 will reduce its clock speed...

... to 1.0 GHz when Cool & Quiet is enabled. Notice the voltage reduction as well. We urge you to use this feature, as it reduces the power draw by more than 40%!

DDR2-800 Memory For Each Processor

Quad FX is the first desktop-type computer that deploys two physical processors that don't communicate via a traditional bus, but rather via HyperTransport (point-to-point). A technology called non-uniform memory access (NUMA) is used, so all processors can access the total physical memory, even though it is managed by two different processors and memory controllers. If application data is located in the other processor's main memory, the coherent HyperTransport link that connects the two CPUs must be used.

In any case, the architecture will cause performance to vary, and will result in somewhat varying benchmark results as well. We measured Windows XP memory performance of around 6 GB/s, while first tests with pre-release versions of Vista look very promising with much higher levels. How much this may impact application performance remains to be seen, though.

5. Moving Socket 1207 Into The Enthusiast Space

AMD's Socket 1207 (also known as Socket F) is a so-called Land Grid Array (LGA) part. Intel has been using this type of socket on its desktop (LGA 775) and server platforms (LGA 771) for some time now, because of its reduced electrical resistance and better reliability at higher operating frequencies.

Unlike with conventional processor sockets, these processors do not have pins. Instead, pin-like connectors can be found within the processor sockets. These are used to "land" the processor on them, instead of inserting pins into the socket.

Nvidia nForce 680a Platform

Technically, the Quad FX platform could be powered by any timely chipset on the market, but AMD and Nvidia had to come up with something substantially improved. Well, here it is: the nForce 680a. It equals the nForce 570 SLI in terms of its northbridge, but nForce 680a uses two of these to provide as many as four x16 PCI Express slots, two powered with 16 lanes, two more running eight lanes, and tons of interfaces. In any case, this provides more than enough bandwidth for a Quad SLI powerhouse and even additional add-on cards. Each of these northbridges is attached to the primary CPU socket via a dedicated X16 HyperTransport link.

Since both nForce 680i components basically are integrated nForce 570 SLI chipsets, users will get twice the connectivity. There can be as many as 20 USB 2.0 ports, 6 + 6 Serial ATA/300 ports with Native Command Queuing (NCQ), four Nvidia Gigabit Ethernet ports and sufficient basic PCI and PCI Express connectivity. Of course, while it's nice to have free resources for future use, this almost looks to be a bit over the top.

6. Asus L1N64-SLI WS Dual-Socket

The first (and still only) motherboard product to support the Quad FX platform and the FX-70 processor series is the L1N64-SLI WS from Asus. It doesn't fully exploit the impressive number of interfaces offered by the nForce 680a chipset, but Asus did an amazing job in squeezing so many components - including two processor sockets - onto a standard ATX motherboard. Let's not forget that typical workstation boards offer a fraction of the L1N64's feature set, or feature larger motherboard dimensions.

The blue PCI Express x16 slots are powered by 16 PCIe lanes each; the black ones are the secondary connectors that run x8 PCIe. There is one more x1 PCIe slot and a single 32-bit PCI slot. Asus implemented two out of four possible Gigabit Ethernet ports; having four would have allowed this platform to act as a software-powered switch!

There are 12 SATA/300 ports on this motherboard. Using 750 GB Seagate Barracuda 7200.10 hard drives this would result in a maximum capacity of 9 TB today! Of course, you would have to look for a suitable case first... There is High Definition Audio, one UltraATA/133 channel and a Firewire controller with up to two ports.

AMD's Reference System

7. Test Setup

System Hardware I
Socket F 2x AMD Athlon 64 FX-74 (Windsor 90 nm, 3.0 GHz, 2x 1 MB L2 Cache)
Motherboard Asus L1N64-SLI WS
Chipset : Nvidia nForce 680a SLI, BIOS : 117, Rev. 1.01
System Hardware II
Socket 775 Intel Core 2 Extreme QX6700 (Kentsfield 64 nm, 2.66 GHz, 2x 4 MB L2 Cache)
Socket 775 Intel Core 2 Extreme EX6800 (Conroe 64 nm, 2.93 GHz, 4 MB L2 Cache)
Motherboard EVGA nForce 680i SLI 775 Mainboard
Chipset : Nvidia nForce 680i SLI, BIOS : P20
Common Hardware
RAM 4x 1024 MB DDR2-1111 (CL 4.0-4-4-12)
Corsair CM2X1024-8888D5 XMS8805v1.2
Graphics Card Nvidia GeForce 7900GTX
GPU : ATI X1900XTX (650 MHz)
RAM : 512 MB GDDR3 (1600 MHz)
Hard Drive I (read) 1x 150 GB 10,000 RPM, 8 MB Cache, SATA/300
Western Digital WD1500ADFD
Hard Drive II (write) 1x 150 GB 10,000 RPM, 8 MB Cache, SATA/300
Western Digital WD1500ADFD
DVD-ROM Gigabyte GO-D1600C (16x)
Software
Nvidia Graphic Forceware 93.71
Nvidia nForce 680i Series Forceware 9.53
DirectX Version : 9.0c (4.09.0000.0904)
OS Windows XP, Build 2600 SP2

Since the test system has a large number of fans - four for the motherboard components, one per processor, two more for the graphics cards, plus the system fans - don’t expect it to be quiet in any way !

Benchmarks And Settings

Benchmarks and Settings
3D-Games
Call Of Duty 2 Version : 1.3 Retail
Video Mode : 1280x960
Anti Aliasing : 4x
Graphics Card : medium
Timedemo demo2
Quake 4 Version : 1.2 (Dual-Core Patch)
Video Mode : 1280x1024
Video Quality : Ultra
THG Timedemo waste.map
timedemo demo8.demo 1 (1 = load textures)
3DMark06 Version 1.0.2
1600x1200 - 32 bit
Anti Aliasing : 4x
Graphics and CPU Default Benchmark
Video
TMPEG 3.0 Express Version : 3.0.4.24 (no Audio)
182 MB VOB MPEG2-source (704x576) 16:9
DivX 6.4 Version : 6.4 (4 Logical CPUs)
Definition Profile : High Definition Profile
Multipass, 3000 kbit/s
Encoding mode : Insane Quality
XviD 1.2.0 Version : 1.2.0 SMP Beta
Encoding type : Twopass - Single pass
Profile @ Level : DXN HT PAL
Target size (kbytes) : 570000
MainConcept H.264 Encoder v2 Version : 2.1
2:19 min MPEG2-source 1920x1080 to H.264
Profile : High
Audio : AAC
Stream : Program
Audio
Lame MP3 Version 3.97 Beta 2 (11-29-2005)
Audio CD "Terminator II SE", 74 min
wave to mp3
160 kbps
OGG Version 1.1.2 (Intel P4 MOD)
Version 1.1.2 (Intel AMD MOD)
Audio CD "Terminator II SE", 74 min
wave to ogg
Quality : 5
Applications
Winrar Version 3.61 (Multi-Core)
(303 MB, 47 Files, 2 Folders)
Compression = Best
Dictionary = 4096 kB
Autodesk 3D Studio Max Version : 8.0
Characters "Dragon_Charater_rig"
rendering HTDV 1920x1080
Adobe Photoshop CS 2 Version : 9.0.1
VT-Runtime Script
Rendering from 5 Pictures (66 MB, 7 Filters)
Cinebench Version 9.5
64 Bit
nCPU, 1 CPU
Synthetic
Everest Version 3.01.652
Cache & Memory Benchmark
PCMark05 Pro Version : 1.1.0
CPU and Memory Tests
Windows Media Player 10.00.00.3646
Windows Media Encoder 9.00.00.2980
SiSoftware Sandra 2007 Version 2007.5.11.17
CPU Test = CPU Arithmetic / MultiMedia
Memory Test = Bandwidth Benchmark
Memory Latency Test = ns
Other
Windows Media Player 10 Version : 10.00.00.36.46
8. Benchmark Results

3D-Games

9. Video

Audio

10. Applications

11. Synthetic

Want more comparison? Check out our Interactive CPU Charts:

12. Power Consumption Tests

In addition to measuring the energy requirements at a specific time, we added another test run to the suite, which measures the real energy consumption after running 3DMark06 for 30 minutes. We used the same 1,000 W power supply unit to ensure a fair test environment.

13. Conclusion

The results of our tests are somewhat ambiguous, so let's first summarize the pros and cons:

Pro AMD Quad FX:

  • Very sophisticated, state-of-the art 32/64-bit uber-enthusiast platform
  • Huge bandwidth between processor(s) and northbridges
  • Dedicated main memory for each processor
  • Works with mainstream DDR2-800 memory
  • Dual-processor system very likely to scale better under Windows Vista Ultimate Edition
  • Entry-level Athlon 64 FX-70 twin-processor kits available at $599
  • Dual-socket platform allows for future upgrade to quad core processors (eight cores per system)

Contra AMD Quad FX:

  • Very high power requirements due to two processors
  • High cooling requirements (air conditioning?)
  • Power draw of 500+ W requires at least a 750 W power supply
  • Very high platform costs (four memory modules and a $400+ motherboard)
  • Platform benefits not noticeable with most of today's applications
  • Virtually no overclocking margins for the FX-74 top model
  • Only a single motherboard available (Asus); likely not available in retail, but via system integrators

Although the primary goal of releasing a workstation-type enthusiast PC for the desktop might have been to answer Intel's quad core offering, AMD has invested a lot in the platform. No other solution offers the bandwidth reserves of a Quad FX, and no other solutions promise such a predictable upgrade. As soon as AMD's "Agena FX" quad cores become available next year you'll be able to replace today's FX-70 series processors and substantially increase your computing power. So, clearly, the platform is brilliant from a technology standpoint.

Unfortunately, considering the effort required to bring this platform into life, the results aren't terribly impressive in terms of both performance and performance per watt. Why should I buy a pair of processors, an expensive and very special motherboard and a heavy duty 1,000 W power supply when I can get almost the same experience with much cheaper mainstream components and Intel's quad core processor? Sure, I won't be able to upgrade an Intel machine to eight cores by the middle of 2007, but AMD's Quad FX powerhouse will be outdated by that time as well. HyperTransport 3.0, with double the clock speed, is waiting for its spring debut, and although the quad core Agena FX will run in today's Quad FX systems, it will have to live with standard HT link speeds.

We still see one reason to go for Quad FX right away: if your ego can live with the FX-70 entry-level model, you will have to pay only a somewhat reasonable $599. Most likely you'll be able to overclock the CPUs to at least 2.8 GHz, likely even to 2.9 or 3.0 GHz, by increasing the base clock speed. If so, you'll get kick-butt performance for a very acceptable price, considering that Intel's Core 2 Extreme QX6700 is still at $999. All you have to decide is whether or not you want to hunt down one of the L1N64-SLI WS motherboards from Asus.

Another possible reason to go for Quad FX instead of Core 2 Quad is the supposedly better scalability of the former under Windows Vista Ultimate Edition. The top version of Microsoft's new operating system will be smarter when it comes to multi-threading and multi-tasking, which could favor AMD's new uber-platform. But again, we cannot say we're particularly convinced that we should spend a lot of money on a motherboard and expectedly $400 on Vista Ultimate in January.

The bottom line is obvious: Quad FX is hot, but its time has not yet come.

Related Article:
Game Over? Core 2 Duo Knocks Out Athlon 64

Join our discussion on this topic