Due to issues with customs/logistics (no fault of AMD’s), we’re a bit late with our Bulldozer review. I suppose that is fitting in this case, but luckily, this gives us the chance to take a bit more time with the review, testing the Bulldozer CPU under its full model lineup, or at least as close as we can get to it.

This is important because, unless you’ve been living under a rock, you have probably heard that the FX 8150 CPU is a huge disappointment. At least that is the sentiment that comes across after reading some reviews, and the reaction on various communities.

Depending on where you go, it’s either “too future oriented” or a “complete unsalvageable failure”. Either way, we already know a few things about the CPU codenamed Zambezi. First, it offers nothing for those looking for the ultimate in performance. AMD didn’t even bother to shoot for the Core i7 2600K or above. Instead, the top-of-the-line FX chip is the MSRP $245 FX-8150, which takes direct aim at the $216 i5-2500K. We’re used to seeing this from AMD, and for a long time the Phenom II was a solid choice for anyone not looking to buy the fastest chip possible, but a solid performer for a good price. I think what got so many people upset was that this was supposed to change with Bulldozer, but it didn’t.

Secondly, it really is ‘too future oriented’ no matter which way you look at it. To understand why, we have to look at how the architecture looks, and how the CPU works.

How Bulldozer Works

First, here is what Bulldozer looks like under the hood:

As you can see, these new processors are built around the “Bulldozer” module, which itself has two x86 integer cores. This is quite different from all previous processors, which cram on several individual processors on one die, and share a large pool of L3 cache. Refer to the Sandy Bridge die map for instance. What Bulldozer does is allow AMD to scale CPUs and cache to a much higher level – this FX 8150 contains 8MB of L2 cache AND 8MB of L3 cache. Compare that to Phenom II X6′s 3MB of L2 and 6MB of L3, and Core i7 2600K’s 1MB of L2 and 6MB of L3. To really understand how Bulldozer is different though, we must go deeper (dun dun dun-duuuun)

This is what each of those four Bulldozer modules look like. There are a pair of Integer cores each with its own L1 cache, but the key here is that there is only one FPU core that is shared between the two. The result of this is what AMD says is a larger, faster, higher clocked FPU with several new instruction sets – including most of the AVX instructions we already saw introduced with Sandy Bridge, and a few more of their own.

This is where we get to the ‘too future oriented’ part. Along with the FP scheduler, you can see the two 128-bit “FMAC” that make up the FPU core. How AMD describes FMA is “(A=B+C*D)” or the ability to add and multiply at the same time. Both AMD and Nvidia GPUs support this style of FPU processing, which led the way for GPU shader accelerated software. As is made evident by such software, this offers a huge benefit for certain scenarios, such as video encoding and rendering. I suppose you could think of this part of the CPU as the first step towards a more “GPU-like” architecture. This is what we’ll be seeing more of in the future – FMA is probably going to be a huge deal. AMD introduced it now, while Intel doesn’t  have it on their roadmap until their next “Tock” microarchitecture change in 2013.

That brings up another issue though – AMD uses a different version (FMA4) from what Intel will use (FMA3). While they are very similar, they will still require separate code to support each. Considering the market share situation, we will have to assume that if a developer is going to support one, it is going to be Intel’s. Therefore, AMD will be forced to switch back to FMA3 on their next CPU, making FMA4 essentially irrelevant.

So AMD has added a feature that everyone knows will eventually be extremely useful, but they are about 2 years too early, and will have to switch around their structure of it. There is some software out there that supports AVX and even AMD’s extra XOP instruction set which they brought over from SSE5 after abandoning it for AVX, so we’ll be able to see what the future holds for Bulldozer, in addition to seeing how it fares with today’s software most people use.

The Platform

Most of the time, when a new CPU is released, there is a new platform to go along with it. However with  Bulldozer, the platform is actually fully backwards compatible with Phenom II chips, and was introduced about six months ago. In fact, it is based on the exact same silicon as the 8xx chipset, and the only difference is a slightly faster HT link between the CPU and North Bridge. One other minor change is based on the CPU and not the chipset, and that is the support for DDR3-1866 memory instead of standard DDR3-1333 on the Phenom II.

So unlike the ‘lower end’ but newer FM1 platform used for Llano CPUs, the 9xx platform doesn’t have native USB 3.0 support. However, what it does have is a ton of PCI-E connectivity, with support for 16+16x lanes of PCI-E for CrossFire. On an Intel Z68 board, the most you’ll get is 8+8x.

Now that we’re properly acquainted with Bulldozer, let’s get on with the review! First, some words about our testing procedure.