Last week Johan posted his thoughts from an server/HPC standpoint on AMD's roadmap. Much of my analysis was limited to desktop/mobile, so if you're making million dollar server decisions then his article is better suited for your needs.

He also unveiled a couple of details about AMD's Bulldozer architecture that I thought I'd call out in greater detail. Johan has been working on a CMP vs. SMT article so I'll try to not step on his toes too much here.

It all started about two weeks ago when I got a request from AMD to have a quick conference call about Bulldozer. I get these sorts of calls for one of two reasons. Either:

1) I did something wrong, or
2) Intel did something wrong.

This time it was the former. I hate when it's the former.

It's called a Module

This is the Bulldozer building block, what AMD is calling a Bulldozer Module:

AMD refers to the module as being two tightly coupled cores, which starts the path of confusing terminology. A few of you wondered how AMD was going to be counting cores in the Bulldozer era; I took your question to AMD via email:

Also, just to confirm, when your roadmap refers to 4 bulldozer cores that is four of these cores:

http://images.anandtech.com/reviews/cpu/amd/FAD2009/2/bulldozer.jpg

Or does each one of those cores count as two? I think it's the former but I just wanted to confirm.

AMD responded:

Anand,

Think of each twin Integer core Bulldozer module as a single unit, so correct.

I took that to mean that my assumption was correct and 4 Bulldozer cores meant 4 Bulldozer modules. It turns out there was a miscommunication and I was wrong. Sorry about that :)

Inside the Bulldozer Module

There are two independent integer cores on a single Bulldozer module. Each one has its own L1 instruction and data cache (thanks Johan), as well as scheduling/reordering logic. AMD is also careful to mention that the integer throughput of one of these integer cores is greater than that of the Phenom II's integer units.

Intel's Core architecture uses a unified scheduler fielding all instructions, whether integer or floating point. AMD's architecture uses independent integer and floating point schedulers. While Bulldozer doubles up on the integer schedulers, there's only a single floating point scheduler in the design.

Behind the FP scheduler are two 128-bit wide FMACs. AMD says that each thread dispatched to the core can take one of the 128-bit FMACs or, if one thread is purely integer, the other can use all of the FP execution resources to itself.

AMD believes that 80%+ of all normal server workloads are purely integer operations. On top of that, the additional integer core on each Bulldozer module doesn't cost much die area. If you took a four module (eight core) Bulldozer CPU and stripped out the additional integer core from each module you would end up with a die that was 95% of the size of the original CPU. The combination of the two made AMD's design decision simple.AMD has come back to us with a clarification: the 5% figure was incorrect. AMD is now stating that the additional core in Bulldozer requires approximately an additional 50% die area. That's less than a complete doubling of die size for two cores, but still much more than something like Hyper Threading.

The New Way to Count Cores

Henceforth AMD is referring to the number of integer cores on a processor when it counts cores. So a quad-core Zambezi is made up of four integer cores, or two Bulldozer modules. An eight-core would be four Bulldozer modules.

A hypothetical quad-core Bulldozer. Presumably the L3 cache would be shared by both modules.

A hypothetical eight-core Bulldozer. Presumably the L3 cache would be shared by all four modules.

It's a distinct shift from AMD's (and Intel's) current method of counting cores. A quad-core Phenom II X4 is literally four Phenom II cores on a single die, if you disabled three you would be left with a single core Phenom II. The same can't be said about a quad-core Bulldozer. The smallest functional block there is a module, which is two cores according to AMD.

Better than Hyper Threading?

Intel doesn't take, at least today, quite aggressive of a step towards multithreading. Nehalem uses SMT to send two threads to a single core, resulting in as much as a 30% increase in performance:

The added die area to enable HT on Nehalem is very small, far less than 5%.

AMD claims that the performance benefit from the second integer core on a single Bulldozer module is up to 80% on threaded code. That's more than what AMD could get through something like Hyper Threading, but as we've recently found out the impact to die size is not negligible. It really boils down to the sorts of workloads AMD will be running on Bulldozer. If they are indeed mostly integer, then the performance per die area will be quite good and the tradeoff worth it. Part of the integer/FP balance does depend on how quickly the world embraces computing on the GPU however...

According to AMD's roadmaps, Zambezi will use either 4 or 8 Bulldozer cores (that's 2 or 4 modules). The quad-core Zambezi should have roughly 10 - 35% better integer performance than a similarly clocked quad-core Phenom II. An eight-core Zambezi will be a threaded monster.

No GPU, for Now

The first APU from AMD will be Llano, but based on existing Phenom II cores. The move to a new manufacturing process combined with the first monolithic CPU/GPU is enough to do at once, there's no need to toss in a brand new microarchitecture at the same time.

AMD did add that eventually, in a matter of 3 - 5 years, most floating point workloads would be moved off of the CPU and onto the GPU. At that point you could even argue against including any sort of FP logic on the "CPU" at all. It's clear that AMD's design direction with Bulldozer is to prepare for that future.

In recent history AMD's architectural decisions have predicted, earlier than Intel, where the the microprocessor industry was headed. The K8 embraced 64-bit computing, a move that Intel eventually echoed some years later. Phenom was first to migrate to the 3 level cache hierarchy that we have today, with private L2 caches. Nehalem mimicked and improved on that philosophy. Bulldozer appears to be similarly ahead of its time, ready for world where heterogenous CPU/GPU computing is commonplace. I wonder if we'll see a similar architecture from Intel in a few years.