The Architecture

We'll start, logically, at the front end of a Bulldozer module. The fetch and decode logic in each module is shared by both integer cores. The role this logic plays is to fetch the next instruction in the thread being executed, decode the x86 instruction into AMD's own internal format, and pass the decoded instruction onto the scheduling hardware for execution.

AMD widened the K8 front end with Bulldozer. Each module is now able to fetch and decode up to four x86 instructions from a single thread in parallel. Each of the four decoders are equally capable. Remembering that each Bulldozer module appears as two cores, the front end can only pick 4 instructions to fetch and decode from a single thread at a time. A single Bulldozer module can switch between threads as often as every clock.

Decode hardware isn't very expensive on its own, but duplicating it four times across multiple cores quickly adds up. Although decode width has increased for a single core, multi-core Bulldozer configurations can actually be at a disadvantage compared to previous AMD architectures. Let's look at the table below to understand why:

Front End Comparison
  AMD Phenom II AMD FX Intel Core i7
Instruction Decode Width 3-wide 4-wide 4-wide
Single Core Peak Decode Rate 3 instructions 4 instructions 4 instructions
Dual Core Peak Decode Rate 6 instructions 4 instructions 8 instructions
Quad Core Peak Decode Rate 12 instructions 8 instructions 16 instructions
Six/Eight Core Peak Decode Rate 18 instructions (6C) 16 instructions 24 instructions (6C)

For a single instruction thread, Bulldozer offers more front end bandwidth than its predecessor. The front end is wider and just as capable so this makes sense. But note what happens when we scale up core count.

Since fetch and decode hardware is shared per module, and AMD counts each module as two cores, given an equivalent number of cores the old Phenom II actually offers a higher peak instruction fetch/decode rate than the FX. The theory is obviously that the situations where you're fetch/decode bound are infrequent enough to justify the sharing of hardware. AMD is correct for the most part. Many instructions can take multiple cycles to decode, and by switching between threads each cycle the pipelined front end hardware can be more efficiently utilized. It's only in unusually bursty situations where the front end can become a limit.

Compared to Intel's Core architecture however, AMD is at a disadvantage here. In the high-end offerings where Intel enables Hyper Threading, AMD has zero advantage as Intel can weave in instructions from two threads every clock. It's compared to the non-HT enabled Core CPUs that the advantage isn't so clear. Intel maintains a higher instantaneous decode bandwidth per clock, however overall decoder utilization could go down as a result of only being able to fill each fetch queue from a single thread.

After the decoders AMD enables certain operations to be fused together and treated as a single operation down the rest of the pipeline. This is similar to what Intel calls micro-ops fusion, a technology first introduced in its Banias CPU in 2003. Compare + branch, test + branch and some other operations can be fused together after decode in Bulldozer—effectively widening the execution back end of the CPU. This wasn't previously possible in Phenom II and obviously helps increase IPC.

A Decoupled Branch Predictor

AMD didn't disclose too much about the configuration of the branch predictor hardware in Bulldozer, but it is quick to point out one significant improvement: the branch predictor is now significantly decoupled from the processor's front end.

The role of the branch predictor is to intercept branch instructions and predict their target address, rather than allowing for tons of cycles to go by until the branch target is known for sure. Branches are predicted based on historical data. The more data you have, and the better your branch predictors are tuned to your workload, the more accurate your predictions can be. Accurate branch prediction is particularly important in architectures with deep pipelines as a mispredict causes more instructions to be flushed out of the pipe. Bulldozer introduces a significantly deeper pipeline than its predecessor (more on this later), and thus branch prediction improvements are necessary.

In both Phenom II and Bulldozer, branches are predicted in the front end of the pipe alongside the fetch hardware. In Phenom II however, any stall in the fetch pipeline (e.g. fetching an instruction that wasn't in cache) would stop the whole pipeline including future branch predictions. Bulldozer decouples the branch prediction hardware from the fetch pipeline by way of a prediction queue. If there's a stall in the fetch pipeline, Bulldozer's branch prediction hardware is allowed to run ahead and continue making future predictions until the prediction queue is full.

We'll get to the effectiveness of this approach shortly.

Scheduling and Execution Improvements

As with Sandy Bridge, AMD migrated to a physical register file architecture with Bulldozer. Data is now only stored in one location (the physical register file) and is tracked via pointers back to the PRF as operations make their way through the execution engine. This is a move to save power as copying data around a chip is hardly power efficient.

The buffers and queues that feed into the execution engines of the chip are all larger on Bulldozer than they were on Phenom II. Larger data structures allows for better instruction level parallelism when trying to execute operations out of order. In other words, the issue hardware in Bulldozer is beefier than its predecessor.

Unfortunately where AMD took one step forward in issue hardware, it does a bit of a shuffle when it comes to execution resources themselves. Let's start with the positive: Bulldozer's integer execution cores.

Integer Execution

Each Bulldozer module features two fully independent integer cores. Each core has its own integer scheduler, register file and 16KB L1 data cache. The integer schedulers are both larger than their counterparts in the Phenom II.

The biggest change here is each integer core now has two ports instead of three. A single integer core features two AGU/ALU ports, compared to three in the previous design. AMD claims the third ALU/AGU pair went mostly unused in Phenom II, and as a result it's been removed from Bulldozer.

With larger structures feeding into the integer cores, AMD should be able to have an easier time of making use of the integer units than in previous designs. AMD could, in theory, execute more integer operations per core in Phenom II however AMD claims the architecture was typically bound elsewhere.

The Shared FP Core

A single Bulldozer module has a single shared FP core for use by up to two threads. If there's only a single FP thread available, it is given full access to the FP execution hardware, otherwise the resources are shared between the two threads.

Compared to a quad-core Phenom II, AMD's eight-core (quad-module) FX sees no drop in floating point execution resources. AMD's architecture has always had independent scheduling for integer and floating point instructions, and we see the same number of execution ports between Phenom II cores and FX modules. Just as is the case with the integer cores, the shared FP core in a Bulldozer module has larger scheduling hardware in front of it than the FPU in Phenom II.

The problem is AMD had to increase the functionality of its FPU with the move to Bulldozer. The Phenom II architecture lacks SSE4 and AVX support, both of which were added in Bulldozer. Furthermore, AMD chose Bulldozer as the architecture to include support for fused multiply-add instructions (FMA). Enabling FMA support also increases the relative die area of the FPU. So while the throughput of Bulldozer's FPU hasn't increased over K8, its capabilities have. Unfortunately this means that peak FP throughput running x87/SSE2/3 workloads remains unchanged compared to the previous generation. Bulldozer will only be faster if newer SSE, AVX or FMA instructions are used, or if its clock speed is significantly higher than Phenom II.

Looking at our Cinebench 11.5 multithreaded workload we see the perfect example of this performance shuffle:

Cinebench 11.5—Multi-Threaded

Despite a 9% higher base clock speed (more if you include turbo core), a 3.6GHz 8-core Bulldozer is only able to outperform a 3.3GHz 6-core Phenom II by less than 2%. Heavily threaded floating point workloads may not see huge gains on Bulldozer compared to their 6-core predecessors.

There's another issue. Bulldozer, at least at launch, won't have to simply outperform its quad-core predecessor. It will need to do better than a six-core Phenom II. In this comparison unfortunately, the Phenom II has the definite throughput advantage. The Phenom II X6 can execute 50% more SSE2/3 and x87 FP instructions than a Bulldozer based FX.

Since the release of the Phenom II X6, AMD's major advantage has been in heavily threaded workloads—particularly floating point workloads thanks to the sheer number of resources available per chip. Bulldozer actually takes a step back in this regard and as a result, you will see some of those same workloads perform worse, if not the same as the outgoing Phenom II X6.

Compared to Sandy Bridge, Bulldozer only has two advantages in FP performance: FMA support and higher 128-bit AVX throughput. There's very little code available today that uses AMD's FMA instruction, while the 128-bit AVX advantage is tangible.

Cache Hierarchy and Memory Subsystem

Each integer core features its own dedicated L1 data cache. The shared FP core sends loads/stores through either of the integer cores, similar to how it works in Phenom II although there are two integer cores to deal with now instead of just one. Bulldozer enables fully out-of-order loads and stores, an improvement over Phenom II putting it on parity with current Intel architectures. The L1 instruction cache is shared by the entire bulldozer module, as is the L2 cache.

The instruction cache is a large 64KB 2-way set associative cache, similar in size to the Phenom II's L1 cache but obviously shared by more "cores". A four-core Phenom II would have 256KB of total L1 I-Cache, while a four core Bulldozer will have half of that. The L1 data caches are also significantly smaller than Bulldozer's predecessor. While Phenom II offered a 64KB L1 D-Cache per core, Bulldozer only offers 16KB per integer core.

The L2 cache is much larger than what we saw in multi-core Phenom II designs however. Each Bulldozer module has a private 2MB L2 cache.

There's a single 8MB L3 cache that's shared among all Bulldozer modules on a chip. In its first incarnation, AMD has no plans to offer a desktop part without an L3 cache. However AMD indicated that the L3 cache was only really useful in server workloads and we might expect future Bulldozer derivatives (ahem, Trinity?) to forgo the L3 cache entirely as a result.

Cache accesses require more clocks in Bulldozer, due to a combination of size and AMD's desire to make Bulldozer a very high clock speed part...

Introduction The Pursuit of Clock Speed
POST A COMMENT

429 Comments

View All Comments

  • Saxie81 - Wednesday, October 12, 2011 - link

    Ouch.... Not looking good. :S

    Thanks for the reply, again great review!!
    Reply
  • velis - Wednesday, October 12, 2011 - link

    Ignoring the power consumption it seems to me that @4.6GHz it should start being quite competitive.
    So can we expect base clocks to rise once significant volume of these chips starts getting out and GloFo refines the process?
    I also must admit I didn't expect 2 bn transistors. All the time AMD was bragging about how much they saved and then we get this behemoth. No wonder they have process issues. Such big chips always do.
    Reply
  • cfaalm - Wednesday, October 12, 2011 - link

    Well it is an 8-core, not a 4 core. 2x 995M (Sandybridge 4C) almost 2B, though I am sure the multply isn't exactly correct. A lot of it depens on the L3/L2 RAM amounts. The savings seem to be minimal.

    I am still confused about why they so deliberately chose to go with a relatively low single thread performance. My main application is multithreaded, but since it's such a mixed bag overall I am pretty unsure if this will be my next CPU, unless I get to see convincing Cubase 6 benchies. For an FX moniker it needs to perform better than this anyway.

    I'll throw in a lyric from The Fixx
    "It doesn't mean much now, it's built for the future."
    Reply
  • TekDemon - Wednesday, October 12, 2011 - link

    Wow, no wonder they say you need water cooling or better to go 5Ghz+. Reply
  • enterco - Wednesday, October 12, 2011 - link

    AMD should send a developer team to CryTek to help them release a patch able to use more cores :) Reply
  • medi01 - Wednesday, October 12, 2011 - link

    Uhm, what about other numbers? Reply
  • IlllI - Wednesday, October 12, 2011 - link

    this might be the final nail in the coffin. We might have to wait longer for it to be competitive? People have literally been waiting for -years- for amd to catch up.
    probably by the time piledriver(or whatever it'll be called) comes out, ib will be out (and even further behind intel)

    btw I think tomshardware tested it with windows 8 and it was still a turd.

    I seriously hope you can get some answers/reasons why amd released such a woeful product. Maybe this was why dirk was fired? All I know is after 7+ years of amd, my next processor will be intel
    Reply
  • Ushio01 - Wednesday, October 12, 2011 - link

    Desktop CPU's are Halo parts and as such are irrelevant. It's the Server and OEM Laptop CPU's were AMD needs to perform and AMD's server share just keeps dropping. Reply
  • lyeoh - Wednesday, October 12, 2011 - link

    Thing is I wouldn't want to use them in my servers: http://us.generation-nt.com/answer/patch-x86-amd-c...

    FWIW when the Athlon64s first came out, we bought a bunch of them, they were not bad, but there were clock issues - the TSCs weren't synchronized. So had to set idle=poll (and thus using more watts).

    You can blame the OS developers, but most people buy new hardware to run existing operating systems and programs on, not future unreleased ones.

    It sure is looking bad for them. I won't be buying AMD CPUs but I hope the fanboys keep them alive ;).
    Reply
  • OCedHrt - Wednesday, October 12, 2011 - link

    "Other than the 8150, only the quad-core FX processors are able to exceed the 3.3GHz clock speed of the Phenom II X6 1100T."

    The 6 core FX is also clocked higher?
    Reply

Log in

Don't have an account? Sign up now