Announcement

Collapse
No announcement yet.

AMD Nearing Full OpenCL 2.0 Support With ROCm 2.0 Compute Stack

Collapse
X
 
  • Filter
  • Time
  • Show
Clear All
new posts

  • #11
    Originally posted by juno View Post
    IIRC OpenCL 2.1/2.2 already use SPIR-V.
    I don't know what OpenCL offers on top of Vulkan. Maybe a generic OpenCL-over-Vulkan driver would be possible, maybe Vulkan will replace OpenCL. No idea.

    But there is problem with SPIR-V in mesa

    Show Mesa progress for the OpenGL, OpenGL ES, Vulkan and OpenCL drivers implementations into an easy to read HTML page.



    Comment


    • #12
      Originally posted by Peter Fodrek View Post


      But there is problem with SPIR-V in mesa

      Show Mesa progress for the OpenGL, OpenGL ES, Vulkan and OpenCL drivers implementations into an easy to read HTML page.


      That only concerns OpenGL SPIR-V implementation. Vulkan drivers (anv and radv) accept SPIR-V since day one, as that is basically the only shader representation available.

      Comment


      • #13
        Originally posted by busukxuan View Post
        Sorry if this is a stupid question, does "SVM" refer to support vector machines? Doesn't sound likely to me that OpenCL would have such a high level primitive.
        Yeah, Shared Virtual Memory. For apps running over KFD/ROCR all system and device memory is in a single virtual address space, and pointers can be shared between CPUs and GPUs. When running over the regular graphics driver each GPU has its own address space separate from CPU address space, but an SVM region can be set up so that CPU and GPU can share pointers within that region.
        Test signature

        Comment


        • #14
          Originally posted by phoronix View Post
          it looks like AMD could be on a strong footing for GPU compute in 2019
          ...depending on what you want to do with it. Not since Hawaii/Grenada, has AMD been competitive in fp64. And Volta did an end-run around Vega's deep learning ambitions, with its Tensor cores.

          I'm not cheerleading for Nvidia, but let's not kid ourselves. AMD seems to have walked away from HPC, and hasn't been competitive in deep learning since Nvidia's GP100 launched, in early 2016. That said, there will certainly be some applications where AMD can be competitive, but Nvidia's new RT cores just edged out another one.

          Comment


          • #15
            @coder: Nvidia started splitting their HPC and GPU lineups, they have been designing discrete top chips specifically for both graphics and compute starting with Pascal. Seems like AMD is doing the same now as they are preparing a HPC chip. It has new DL instructions, so maybe even specific hardware units similar to tensor cores. It has 1/2 FP64 rate, so it should reach Volta if they can manage to get past 1.8 GHz with a reasonable power budget, which I think is possible as they use TSMC's 7nm node and Vega10 already reaches 1.7 GHz.

            I also think that RT has been a topic in the industry for a long time. Even Imagination Technologies managed to merge RT cores in their traditional GPU design. I think it might even be possible for AMD to get it inside Navi. At least, I hope so because I agree that dedicated hardware for RT is great and think it's the future of real time graphics as well. And it also helps very much in professional workloads.

            I think their GPU team is still competitive. They just can't afford to release as many GPUs as Nvidia. As a reminder, in the past few years Nvidia has released GM200, GM204, GM206, GM207, GM208, GP100, GP102, GP104, GP106, GP107, GP108, GV100, and soon GT102, GT104, GT106. In the same timeframe, AMD has only released Tonga, Fiji, Polaris10, 11, 12, Vega10 and soon Vega20. And of course their CPU, APU and semi-custom SoCs. I think it's remarkable that they can keep up to this point and not being left behind much further.

            Comment


            • #16
              My layman's impression is that the big money in machine learning hardware, will initially be in large scale installations.

              I am skeptical that cuda is the barrier to market entry it is credited with for AMD.

              Writing some code doesnt bother those guys much if there is better perf/$ to be had from AMD.

              The amd value proposition works quite well at scale. They dont mind much if it takes 3 AMD gpuS to do the work of 2 high priced nvidias, so long as its cheaper.

              A recent game changer for AMD platform resources, is the explosion in core/thread numbers. That a ~humble TR can simply double cores using the same platform, seems a trick (~give each ccx a slave ccx on an ~easily modified die) that can be repeated on am4 and x399 epyc. Maybe AI will be a little less gpu centric now?

              AMDs crushing dominance in most metrics, must make choosing a sibling amd gpu very compelling for epyc adopters

              Comment


              • #17
                Originally posted by coder View Post
                ... let's not kid ourselves. AMD seems to have walked away from HPC, and hasn't been competitive in deep learning since Nvidia's GP100 launched, in early 2016. That said, there will certainly be some applications where AMD can be competitive, but Nvidia's new RT cores just edged out another one.
                There is never just one way to do things with code.

                a champion team beats a team of champions.

                would u say the same for multi 7nm Vega on a 64 core/128T Epyc in a tightly integrated ecosystem?

                Which btw, has 128 lanes for scads of all important nvme.

                IMO, using nand cleverly as pseudo ram, will be the key to the insatiable memory demands of vast data sets of big end of town AI.

                Coders just use the resources available. If resource options change and improve - who knows what approach will yield the best AI perf/$? Its still a young field.
                Last edited by msroadkill612; 11 September 2018, 06:45 AM.

                Comment


                • #18
                  Originally posted by msroadkill612 View Post
                  There is never just one way to do things with code.
                  Within the scope of a given problem domain, you reach certain fundamental limits. Sure, it's often possible to devise some innovative approach, like using 8-bit integer arithmetic for CNNs. However, so far as we've seen, AMD has merely been following Nvidia's path in deep learning, rather than out-innovating them.

                  Originally posted by msroadkill612 View Post
                  a champion team beats a team of champions.
                  What are you even talking about?

                  Originally posted by msroadkill612 View Post
                  would u say the same for multi 7nm Vega on a 64 core/128T Epyc in a tightly integrated ecosystem?
                  You cut out part of my message. I didn't say it wouldn't compete in any scenario. Whether such a solution is competitive depends on what you're trying to do with it.

                  Comment


                  • #19
                    Originally posted by msroadkill612 View Post
                    I am skeptical that cuda is the barrier to market entry it is credited with for AMD.
                    It's not. CUDA is a detail. The key is to have accelerated support for popular frameworks.

                    The other thing is to make sure those frameworks & surrounding toolchains are competitive with TensorRT, which is to deep learning what CUDA is to parallel programming APIs.

                    Originally posted by msroadkill612 View Post
                    Writing some code doesnt bother those guys much if there is better perf/$ to be had from AMD.
                    I dunno. I think a lot of deep learning practitioners don't want to get bogged down in backend stuff. That said, you might get Google to help AMD optimize TensorFlow, or get FB to help them optimize Caffe2... except that both are now increasingly invested in their own hardware. But I think it shouldn't be too hard for them to get caught up with framework support, whether using their internal developers or perhaps even paying some grad students to help out.

                    The real problem is the numbers. AMD just has no answer for Nvidia's Tensor cores.

                    Originally posted by msroadkill612 View Post
                    The amd value proposition works quite well at scale. They dont mind much if it takes 3 AMD gpuS to do the work of 2 high priced nvidias, so long as its cheaper.
                    Is this hypothetical, or are you actually paraphrasing someone in the real world? Because TFLOPS/W (or TOPS/W) is crucial for anyone using deep learning at scale, and that's where Nvidia pwns AMD.

                    Originally posted by msroadkill612 View Post
                    A recent game changer for AMD platform resources, is the explosion in core/thread numbers. That a ~humble TR can simply double cores using the same platform, seems a trick (~give each ccx a slave ccx on an ~easily modified die) that can be repeated on am4 and x399 epyc. Maybe AI will be a little less gpu centric now?
                    No. You just need to look at the numbers. AMD didn't even bother with AVX-512 support, since they realized it's trying to do something with CPUs that's a much better fit for GPUs.

                    Comment


                    • #20
                      Originally posted by juno View Post
                      Nvidia started splitting their HPC and GPU lineups, they have been designing discrete top chips specifically for both graphics and compute starting with Pascal.
                      That's not exactly clear, since graphics versions of the P100 and V100 have both been released. So, we don't know whether there are actually two versions of the silicon, or if the server versions simply don't use the graphics units. I've read some hints which suggest it's actually the same silicon.

                      Originally posted by juno View Post
                      Seems like AMD is doing the same now as they are preparing a HPC chip. It has new DL instructions, so maybe even specific hardware units similar to tensor cores.
                      They might, but AMD hasn't said as much. The performance numbers they have revealed didn't suggest anything like that.

                      Originally posted by juno View Post
                      It has 1/2 FP64 rate,
                      I don't recall hearing anything about that, but it would be great for them.

                      Originally posted by juno View Post
                      I think it might even be possible for AMD to get it inside Navi. At least, I hope so because I agree that dedicated hardware for RT is great and think it's the future of real time graphics as well. And it also helps very much in professional workloads.
                      I sure hope AMD saw this one coming.

                      I'm still left wondering where, exactly, they would put the RT cores. Their GPUs get less graphics performance per mm^2 than Nvidia's. Unless that situation changes, I don't see how they could justify burning extra silicon on RT hardware. They can't afford to release something that's unable to be cost-competitive with Nvidia. I think that's already happened, with Vega. Its price hasn't dropped to match current GeForce pricing. [checks current prices] ...although it might be getting close.

                      Originally posted by juno View Post
                      They just can't afford to release as many GPUs as Nvidia. As a reminder, in the past few years Nvidia has released GM200, GM204, GM206, GM207, GM208, GP100, GP102, GP104, GP106, GP107, GP108, GV100, and soon GT102, GT104, GT106. In the same timeframe, AMD has only released Tonga, Fiji, Polaris10, 11, 12, Vega10 and soon Vega20.
                      GT106? I hadn't heard about that. According to this, RTX 2070 uses basically the same silicon as RTX 2080.



                      Also, you forgot Polaris20.

                      I think it will be painful for AMD not to release a new GPU, in 2018. The crypto-boom probably saved their bacon.

                      Comment

                      Working...
                      X