Fused Multiply-Add and HSA

Article: AMD's Analyst Update
By: David Kanter (dkanter.delete@this.realworldtech.com), February 7, 2012 11:54 pm
Room: Moderated Discussions
Richard H. (no-spam-please@gmail.com) on 2/7/12 wrote:
---------------------------
>David Kanter wrote:
>"Software that adopts FMA can theoretically double compute performance"
>
>Please explain why.

This is in the context of AMD's Bulldozer core, I'd recommend reading:

http://realworldtech.com/page.cfm?ArticleID=RWT082610181333&p=7

>On Intel's processors, a SIMD multiply instruction plus a
>SIMD add instruction can be completed every clock. I don't understand how a Fused
>Multiply-Add (FMA) increases peak performance at all compared to what Intel already
>does (not even counting 128-bit SIMD vs 256-bit SIMD).

Again, this about AMD not Intel. Bulldozer has 2x128b FMA pipelines, but they are FMA4. Piledriver is FMA3 (which people will actually use as it is compatible with Haswell). So when you don't use an FMA, you leave half the FLOP/s on the table.

>I'd like to hear people's opinions about AMD's Heterogeneous System Architecture
>(HSA). I am struggling to understand how it is possible to have an ISA-agnostic
>interface (therefore needing a Just-In-Time compiler) >without paying a big performance
>penalty.

LLVM manages quite nicely. As GPUs become more similar to one another, it will be easier to write code that can function well on all architectures.

Also, it may be possible to compile at install time (rather than JITing).

>The whole purpose of heterogeneous computing is to get more >performance
>so why would it make sense to use a Just-In-Time compiler?

JITs may not be required. Even if they are, some JITed languages are quite fast.

>It seems like the vast majority of software vendors that make use of GPUs for computing
>only support Nvidia GPUs (for example, Adobe Photoshop, >Adobe Premiere and Neat
>Video). What does HSA provide that could change this >situation?

Because CUDA is on it's way out. It's a proprietary interface for a company with relatively low market share. Software developers would much rather have something that runs on AMD, Intel and Nvidia GPUs.

Intel can get away with proprietary interfaces, because they have massive market share. The rest of the world really cannot (in the context of PCs). Even there, it's not clear which way Intel will head...OpenCL may be looking very attractive for developers.

David
< Previous Post in ThreadNext Post in Thread >
TopicPosted ByDate
Fused Multiply-Add and HSARichard H.2012/02/07 07:29 PM
  Fused Multiply-Add and HSADavid Kanter2012/02/07 11:54 PM
    "CUDA is on its way out"pk2012/02/09 05:37 AM
      "CUDA is on its way out"David Kanter2012/02/09 11:49 AM
        "CUDA is on its way out"Bryan Catanzaro2012/02/09 12:44 PM
          "CUDA is on its way out"David Kanter2012/02/09 05:48 PM
            "CUDA is on its way out"Anon2012/02/09 07:11 PM
              "CUDA is on its way out"David Kanter2012/02/10 12:00 AM
                "CUDA is on its way out"Michael S2012/02/10 04:54 AM
                "CUDA is on its way out"Anon2012/02/10 06:42 PM
                  Proprietary SW interfaces and hardware vendorsDavid Kanter2012/02/10 09:07 PM
                    Proprietary SW interfaces and hardware vendorsAnon2012/02/11 04:08 AM
                      Proprietary SW interfaces and hardware vendorsAnon2012/02/11 04:12 AM
                      Proprietary SW interfaces and hardware vendorsDel2012/02/11 08:39 AM
                  "CUDA is on its way out"Del2012/02/11 08:33 AM
                "NVIDIA is on its way out"Brendan2012/02/10 10:42 PM
                  "NVIDIA is on its way out"anonymous2012/02/11 11:37 AM
              "CUDA is on its way out"Del2012/02/11 08:22 AM
                "CUDA is on its way out"Anon2012/02/11 04:11 PM
                  ispc on GPUsMatt Pharr2012/02/12 10:02 AM
                    ispc on GPUsAnon2012/02/13 03:34 PM
          "CUDA is on its way out"Erwin Coumans2012/02/09 09:39 PM
            "CUDA is on its way out"Brett2012/02/10 12:54 PM
              "CUDA is on its way out"David Kanter2012/02/10 04:00 PM
              "CUDA is on its way out"Anon2012/02/10 06:45 PM
                "CUDA is on its way out"Brett2012/02/10 08:20 PM
                  "CUDA is on its way out"David Kanter2012/02/10 08:58 PM
                    "CUDA is on its way out"Brett2012/02/11 07:15 AM
                      Macroscalarnone2012/02/11 08:40 AM
        "CUDA is on its way out"Anon2012/02/10 06:39 PM
          "CUDA is on its way out"David Kanter2012/02/10 09:19 PM
            "CUDA is on its way out"Anon2012/02/11 04:33 AM
          "CUDA is on its way out"NN2012/02/14 04:24 PM
            "CUDA is on its way out"Anon2012/02/15 04:16 PM
        "CUDA is on its way out"bakaneko2012/02/11 05:31 AM
      "CUDA is on its way out"Aaron Spink2012/02/10 12:29 AM
  Fused Multiply-Add and HSAEduardoS2012/02/08 03:11 PM
    Fused Multiply-Add and HSAFoo_2012/02/09 06:24 AM
Reply to this Topic
Name:
Email:
Topic:
Body: No Text
How do you spell green?