It has been known by many names: Zambezi, Bulldozer, and Orochi. Today we know it as AMD FX, and after six long years of waiting, it’s here.
AMD calls it the future of computing. It’s clear, however, that what they might be saying instead is “Welcome to the future of gaming.” As one Icrontic reporter wrote back in 2010: “2011 will be more exciting than ever” regarding the competition between Intel and AMD. To say that Bulldozer has been hotly anticipated is an understatement. Speculation, leaks, rumors, and the hopes and dreams of PC gamers and overclockers have been running rampant all over the web.
AMD’s answer to this ravenous horde is the FX line of CPUs.
The FX CPU is the third and final component of AMD’s Scorpius platform. The other components are a motherboard with a 900-series chipset and an AMD Radeon HD 6000-series GPU (Read: Icrontic’s review of the SAPPHIRE Radeon HD 6950).
Also available as a cooling option is OEM water cooling—a first for any consumer CPU. While pricing isn’t yet available, it’s expected to add $80-100 to the total cost. We did receive a sample watercooler unit, but not in time for this review. We’ll follow up soon.
The AMD FX series is broken down into three core counts: 4, 6, and 8. The eight core models, at launch, come at two factory clock speeds. Models available at launch time:
Bulldozer technical details
To understand the new core and module layout, it might be easier to have a simplified picture of Intel’s Hyperthreading in mind. Hyperthreading allows two threads to run on a single core simultaneously. Each core is seen as two virtualized cores with only a very small amount of the core’s internals actually duplicated (just enough to store each virtual processor’s state), meaning everything shares a single execution unit. Most of the time it works out, but under the right conditions, the whole thing can get jammed up waiting for instructions to finish.
AMD’s approach is similar, but the key difference is that it’s all done in hardware—nothing is virtualized. Each AMD FX CPU is made up of between one and four modules. Each module contains two full cores… almost. In a traditional CPU, for example the Phenom II, each core would contain its own floating point (FP) scheduler and a pair of 128-bit fused multiply-accumulate (FMAC) units. The Bulldozer modules share a single FP scheduler and pair of 128-bit FMACs, while retaining individual copies of everything else. Multi-threaded software gets real hardware for each thread, and single-threaded software gets full access to the shared resources.
Power management is improved as well. The new core/module layout detailed above is designed to be efficient through use of a shared FP scheduler. New power states have been added, and core speeds can be controlled individually. TurboCORE (read: explanation of TurboCORE) now has settings for all cores instead of just “some” cores (The TurboCORE implementation in 6-core Phenom II CPUs required that three cores lay dormant). Each module can be turned on and off independently, allowing only the needed cores to receive power.
The memory controller has been updated. Even though it’s still limited to dual channel operation (Intel will have quad channel in Ivy Bridge), supported speeds have been bumped to DDR3-1866, and a maximum of 32GB of RAM is supported, providing you can afford 8GB DDR3 modules. Voltage requirements have been lowered to 1.5V. Higher voltage RAM may work, but your mileage may vary.
Several new instructions have been added to the mix as well. Those instructions are AES, AVX, FMA4, XOP, and SSE 4.2. AES will speed encryption of individual files as well as entire disks—programs such as TrueCrypt will benefit tremendously. AVX boosts performance of floating point applications. FMA4 and XOP are used for complex math in applications like OCL Perf Mandelbrot.
Of course, many of the new features can’t be utilized with today’s software, but this is not uncommon. The same happened with the Intel Sandy Bridge CPU’s new features—software must be built to use new features and compilers must be updated to recognize the new instruction sets. This isn’t to say that FX performs poorly with today’s applications—it just won’t fully shine until the tools are updated and developers take advantage of the new features. One of those hurdles is soon to be corrected; Microsoft Visual Studio 2010 SP1 and Open64 v4.2.5.2 support all of the new instructions supported by FX processors.
Speaking of the importance of support, something that must be talked about is the not-so-secret issue with the Windows 7 scheduler that negatively impacts FX processors. Windows 7 is unaware of the shared nature of the FX processor modules. As a result, there is a possibility that opportunities for resource sharing or higher TurboCORE frequencies may go missed. This has been addressed in Windows 8, and can be seen in the Developer Preview.
Okay, you can wake up now. Let’s get to the stuff.
Configuration
Our test system uses the ASUS Crosshair V Formula motherboard, which is AMD’s reference board for this platform. In the limited amount of time available for testing, it’s been a very solid motherboard.
- CPU: AMD FX-8150
- cores: 8
- Speed: 3.6GHz
- Turbo Core (8/2 core): 3.9GHz/4.2GHz
- Cache (L2/L3): 8MB/8MB
- TDP: 125W
- Motherboard: ASUS Crosshair V Formula
- RAM: 16GB DDR3-1600 (9-9-9-24, 1.5V)
- GPU: AMD Radeon HD 6950 2GB
- HDD: OCZ Vertex 2 120GB
- PSU: Thermaltake ToughPower 1200W
Synthetic Benchmarks
PCMark 7
PCMark 7 shows a fairly close race between the FX-8150 and the i7 2600K. In fact, where the numbers aren’t a virtual tie, the two CPUs trade the top position.
3DMark 11
3DMark 11 demonstrates a virtual tie again in most of the individual tests. Intel holds the lead for physics, AMD leads in one of the graphics tests. Everything else is virtually identical.
SiSoft Sandra CPU
Intel holds a lead in almost every aspect of the CPU tests. What’s interesting is the multimedia integer performance, heavily favoring the FX-8150 due to having twice the number of actual integer units.
SiSoft Sandra Memory
The memory controller for the FX-8150 is definitely improved over the previous generation. It’s not quite as fast as the i7 2600K, but it’s a move in the right direction.
wPrime 32M
Remember that thing about the Windows 7 scheduler not playing nicely with FX processors? I suspect this could be the case for wPrime. Regardless, even the older Phenom II X6 1100T outperforms the FX-8150.
Cinebench R11.5
CPU performance for Cinebench again shows Intel in the lead.
DirectCompute
Intel holds the lead in DirectCompute CPU performance.
Applications and real use tests
H.264
Moving along into real applications, media conversion demonstrates close performance between the FX-8150 and i7 2600K. It should also be noted that this is a stock version of the x264 transcoding utility. There is a new version available that takes advantage of the new instructions supported by the FX processors.
MP3 encoding
This single-thread MP3 encoder definitely doesn’t play nicely with the FX-8150.
Gaming benchmarks
Unigine Heaven
This DirectX 11 benchmark is about as close to a tie as you can get. The FX-8150 shows the slightest of leads, but it’s nothing noticeable.
Alien Vs Predator
There’s so little performance difference here that nobody would notice. Less than one frame per second separates all of the CPUs.
H.A.W.X. II
Interestingly, the higher the settings within a resolution, the more the FX-8150 is favored. It’s especially pronounced at 1680×1050 where the FX-8150 gains a 16fps lead over the i7 2600K.
Dawn of War II
Intel holds a huge lead over AMD here, sometimes nearly doubling the FX-8150’s performance.
Batman: Arkham Asylum
The FX-8150 starts out behind the i7 2600K, but with each increase in quality the gap is closed.
Lost Planet 2
With very few exceptions, the FX-8150 holds the lead.
Metro 2033
The FX-8150 holds a performance lead over the i7 2600K at almost every setting. Sometimes it’s just a fraction of a frame, sometimes the gap widens to five frames per second.
How Bulldozer stacks up on the builder side
TDP
The FX-8150 is a 125W part, so it uses more power than the Core i7 2600K, but stopping there would be telling an incomplete story. Looking at the Phenom II X6 1100T we see a six core 95W processor. The Core i7 2600K is a 95W part with four cores. On a watts-per-core basis, the FX-8150 is the most efficient at 15.625W/core, the Phenom II X6 1100T is next at 15.83W/core, while the Core i7 2600K comes in last at 23.75W/core. Move down to the FX-8120 and a 95W part becomes available as well, using a significantly smaller 11.875W/core.
A stock AMD cooler wasn’t available for testing, but the design hasn’t changed from the Phenom II. As previously mentioned, an FX-branded water cooling kit is also available. We’ll be looking at that in the near future as well.
Overclocking
We’ll be exploring the overclocking potential of Bulldozer in a follow-up article. According to what we’ve seen on the web, AMD’s expectations are quite high—Bulldozer was recently used to break the world overclocking record. At a recent event, AMD got FX-8150 overclocks to record setting speeds of 8.4GHz using two cores and some liquid helium. But for us mere mortals, a more modest 4.6GHz can be expected using all eight cores with the stock heatsink. Dropping to two cores could result in speeds as high as 5.0GHz. Got water cooling? You could see a slightly higher 4.9GHz or 5.2GHz for eight or two cores, respectively.
Price
The air cooled version of the FX-8150 retails for $245, although retailers such as TigerDirect are marking them up slightly. The water cooled version should come in around $80-100 higher. For a CPU that plays games as well as the $314.99 Core i7 2600K it’s quite a good deal. Throw in the advertised overclocking and you’ve got yourself one good platform, and this is only the beginning. Things should scale up very nicely from here. The other Bulldozer CPUs are priced as follows: AMD FX-8120—$205, AMD FX-6100—$165, and AMD FX-4100—$115.
Conclusions
Bulldozer is the culmination of five years of effort and considerable heartache for AMD. They had a solid lead on Intel with the Athlon 64, but steadily lost momentum to Intel’s aggressive Tick Tock strategy over the ensuing half decade. AMD seemed to concede the fight for the performance crown and settled for differentiating with a compelling price/performance ratio. In addition, the 2006 acquisition of ATI Technologies bore fruit in the form of the Llano and Ontario APUs—CPUs that integrate graphics silicon. There are some problems, to be sure—yields are low(ish), per-core IPC is lower than it could be, and it isn’t quite the home run AMD needed it to be. But, it’s plenty fast, and it already overclocks like crazy on a very young 32nm process—so there’s definitely room to grow.
So what’s their game? Do they intend for FX to reclaim the performance crown and compete squarely with Sandy Bridge on the performance side? Not precisely. FX is a different kind of processor.
To be perfectly clear: the FX CPUs represent AMD’s vision of the future of computing, not the present. As such, much of the software available today doesn’t run as well as it does on other CPUs. But to leave it at that would only be telling half the story. It’s a solid starting point for future models, it’s the budget champion by a clear margin, and a fair number of its problems can be traced to software. Software patches will change the story for many of the applications that matter; especially as developers become more comfortable with the model and with the new instruction sets.
That being said, gaming performance is definitely a strong point of the FX line—a refreshing change of pace. With one exception in our test suite, the FX-8150, at a minimum, trades the top spot with a Core i7 2600K. Combine that gaming performance with a price significantly lower than the top Sandy Bridge CPU and you have a compelling reason to give the FX-8150 a hard look. Besides that, given the known overclockability of the FX CPUs, it’s not out of the realm of possibility that the FX-8150 could be made to outperform anything out there.
It’s an exciting time to be a PC gamer, and things are only getting better.