8 ms·

With Intel stuck in process hell for a while longer, I primarily hope AMD can gain ground on Nvidia. They are pushing a lot of innovation and gaining market dom

With Intel stuck in process hell for a while longer, I primarily hope AMD can gain ground on Nvidia. They are pushing a lot of innovation and gaining market dominance, both in the ML/GPGPU space and gaming (with ray tracing, upscaling, etc). We need some competition to keep prices low and innovation high medium to long term.

jonplackett 6y ago
Funny I just posted almost the exact same. Seems like we all have a soft spot for AMD!
freeone3000 6y ago
What's their CUDA competitor this year? Are they pushing out client libs for Vulkan or are we sticking with trying to make ROCm work?
- robert_foss 6y ago
  The industry at large is moving to SYCL, which can be implemented on top of OpenCL / CPUs / FPGAs.
  - roel_v 6y ago
    I'm not sure what part of 'the industry' you're referring to, but last time I looked (which was admittedly more than a year ago), there is only one serious SYCL implementation, it's proprietary and the only people pushing SYCL are... drum roll... the people selling that implementation. That's not even considering that OpenCL is in most cases (way) behind CUDA in performance, not to mention that the tooling for OpenCL is outright primitive. And all of that is not even mentioning that nvidia drags their feet massively in supporting newer OpenCL version on their GPU's. Maybe things have changed massively very recently, but I came back from iwocl (the OpenCL and SYCL conference) two years ago massively disappointed. And I still don't see anyone in the slice of the HPC space I'm familiar with (environmental & GIS-related modeling) using it, for a reason I presume.
    - rodburns 6y ago
      Intel are supporting SYCL in a big way. They are implementing SYCL in the open source LLVM project, the project is called DPC++. Argonne National Labs are also using SYCL to program the exascale supercomputer (Aurora) they are building. Whilst it's not the best solution for everyone there is certainly industry support for it. https://software.intel.com/content/www/us/en/develop/tools/oneapi/components/dpc-compiler.html https://software.intel.com/content/www/us/en/develop/tools/o... https://github.com/intel/llvm/blob/sycl/sycl/doc/GetStartedGuide.md https://github.com/intel/llvm/blob/sycl/sycl/doc/GetStartedG... https://www.alcf.anl.gov/support-center/training-assets/roadmap-sycldpc-aurora https://www.alcf.anl.gov/support-center/training-assets/road...
  - paulsmal 6y ago
    this is nice, do you have any more details?
  - freeone3000 6y ago
    Oh, the third option, with absolutely no library support. Or, looking at Khronos's page, even device support. At least OpenCL has the benefit of wide compatibility, even if its performance meant it was unsuitable. Nvidia has settled on CUDA, and has maintained 13 years of compatibility. If SYCL is going to ever compete with CUDA, it's going to need to stick around long enough for people to build on it.
    - throwaway56034 6y ago
      Well, Nvidia invented CUDA, it's no surprise that they won't give up their competitive advantage by properly supporting OpenCL or SYCL. Edit: Also, SYCL should be supported by most things already supporting OpenCL, like Tensorflow for example.
    - fortran77 6y ago
      My firm does GPGPU programming for audio processing. Nobody in this area even thinks of using anything other than CUDA. It has far more support, tools, compilers, etc.
      - Aaargh20318 6y ago
        What about the Mac side of things ? Mac's are still pretty big in audio production, but they don't support CUDA (and don't come with Nvidia GPUs any more anyway). How does Metal Compute stack up ?
        fortran77 6y ago
        We don't use Macs. While many uusicians use macs, the people on the engineering/scientific side of things don't. We're not doing music audio.
  - dragontamer 6y ago
    I think OpenMP device offload has a better chance than SYCL or OpenCL. But I'm not in the industry. CUDA's single-source environment is very good compared to SYCL or OpenCL. OpenMP still has single-source (just #pragma target), and is already adopted by the HPC crowd for multithreaded programming. NVidia and AMD support seems to be growing, especially because OpenMP is important to the supercomputer folks. (Summit and Frontier) ROCm is CUDA-like and mostly works. The main issue is that CUDA will always remain a few steps ahead as AMD is forced to play "catchup". (ex: Cooperative groups are useful but unimplemented for now in ROCm). ROCm's main issue is being stuck in Linux, they need some Windows support if they want ROCm to really take off.
  - shmerl 6y ago
    Didn't they plan to merge OpenCL and Vulkan?
- arnon 6y ago
  ROCm and HIP - https://github.com/ROCm-Developer-Tools/HIP https://github.com/ROCm-Developer-Tools/HIP, as it has been for the last 3-4 years
  - rrss 6y ago
    unless you have navi, in which case rocm is not supported and doesn't work and I guess AMD expects you to use opencl.
    - dragontamer 6y ago
      Navi just began to compile a week or two ago. Still 'unsupported', but its slowly gaining features. ROCm is a bit uneven with what cards it supports. Ex: Rx 550 never was supported (even though Polaris, the rest of the 5xx series worked). It seems like the only cards ROCm works for are the ones that share a chip and/or driver with AMD's "Machine Intelligence" line of cards. (MI50, MI60, etc. etc.). Which are Fiji (Rx Fury), Polaris (Rx 580), Vega, and Radeon VII. CUDA is a "works on all NVidia devices" kind of thing. So I think people are surprised when they have to read the docs for AMD's ROCm.
      - numpad0 6y ago
        Last time I tried, PCIe 3.0 was also a requirement, although that wasn't much of an issue unless it was some nonstandard configuration like an RX 580 in a junkyard find A4 APU machine.
      - Vvector 6y ago
        > Navi just began to compile a week or two ago. Still 'unsupported', but its slowly gaining features. 14 months after release, still unsupported. AMD needs to step up to the plate if they want to be seriously considered in the AI world.
        dragontamer 6y ago
        Its pretty clear what's going on: AMD doesn't have the resources to support all the GPUs they release on ROCm. They focus on the "MI" line of cards, which have similarities to the consumer GPUs. (Ex: Rx Fury is very similar to MI8). Because MI8 and Rx Fury have similarities, AMD gives ROCm support to both. Ditto with MI25 and Vega, and MI6 and Rx 580. EDIT: With "Navi" gaining features in ROCm repos, my bet is that a new MI-card based on RDNA2 might be coming out soon. Maybe its Arcturus (the Exascale GPU for the Frontier supercomputer).
reitzensteinm 6y ago
Chips are in the pipeline for many years, so products being released today are probably on the tail end of being impacted by AMD cutting GPU budgets to the bone to focus on Zen. I wouldn't expect miracles. Given AMD's market cap and financial success, I bet they're working on competitive designs now. There's no money in being an also ran in the space.
- DisjointedHunt 6y ago
  The information available around Big Navi (The GPUs expected to launch next month) put them at a +50% performance jump from the last generation and likely able to compete with the Nvidia flagship 3080: https://www.tomshardware.com/news/amd-big_navi-rdna2-all-we-know https://www.tomshardware.com/news/amd-big_navi-rdna2-all-we-... I'd really urge anyone wanting to opine on these issues to watch the official analyst day videos and read through the releases AMD has put out so far in terms of architecture and other tech decisions they've made: https://www.youtube.com/watch?v=LMNwLVJQzGs&t=10064s https://www.youtube.com/watch?v=LMNwLVJQzGs&t=10064s
  - mbell 6y ago
    > The information available around Big Navi (The GPUs expected to launch next month) put them at a +50% performance jump from the last generation and likely able to compete with the Nvidia flagship 3080 I guess it depends on what benchmark you care about but for AMD to compete with a 3080, which is 70-100% faster than a 2080, which is 10-20% faster than AMD's current flagship, they need to do a lot better than +50%.
    - strictnein 6y ago
      Some of the leaked benchmarks for the 3080 put it more in the range of 30-50% better than the 2080 series: "30% Faster Than The RTX 2080 Ti & 50% Faster Than the RTX 2080 SUPER on Average" https://wccftech.com/nvidia-geforce-rtx-3080-graphics-card-benchmarks-game-performance-leak/ https://wccftech.com/nvidia-geforce-rtx-3080-graphics-card-b...
deeeeeplearning 6y ago
>gaining market dominance, both in the ML/GPGPU space Is this really the case? Last I checked all the major Deep Learning toolsets ran off CUDA/cuDNN and there was nothing comparable for AMD hardware.
- travisporter 6y ago
  I agree. I think OP was referring to nvidia?
- d0mdo0ss 6y ago
  anyone have some experience on this? I'm in the market for an AMD laptop capable of running standard sci-kit/pytorch/etc but these seem optimized for NVIDIA cards. I'm curious about the outlook for these trending AMD cards.
  - jandrese 6y ago
    Generally if you want to do ML tasks you want nVidia. They put the work in early to build the tooling so now the default assumption is that you're on nVidia hardware. It is possible to do some stuff on AMD cards, but you'll be on the cutting edge for that platform re-solving problems that were already solved on the nVidia side.
    - d0mdo0ss 6y ago
      yes that was my conclusion after some research on this. there is an open issue on github for AMD support on pytorch, and looks like something works on arch linux, but really sounds like support is still in the hacking stage and far from production mode. https://github.com/pytorch/pytorch/issues/10657 https://github.com/pytorch/pytorch/issues/10657
      - t-vi 6y ago
        Support for PyTorch on ROCm is fairly good. I have built it from the dev branch without much trouble for all year. There has been ROCm CI even longer, my patch to print ROCm system information for bug reports was merged last week. If you know where to look (i.e. it's public but unannounced) you can see that nightly wheels have been built for the last few days. So I would expect that some time between now and the Developer Day in November we'll see ROCm appear on PyTorch's "get started" page.
marmaduke 6y ago
I recently ported a code to AVX512 (with ISPC) and was blown away by the performance. Sure, maybe fma can cause throttling but for me, I could replace a 28 core Haswell machine by a 10 core Skylake X (Xeon W specifically) with the new instruction set. A fatter one (16 core?) would compete with a high end Nvidia card and maybe a 32 core EPYC. Sure, it’s just my benchmark, but it is competitive.

jonplackett 6y ago

Funny I just posted almost the exact same. Seems like we all have a soft spot for AMD!

freeone3000 6y ago

What's their CUDA competitor this year? Are they pushing out client libs for Vulkan or are we sticking with trying to make ROCm work?

robert_foss 6y ago
The industry at large is moving to SYCL, which can be implemented on top of OpenCL / CPUs / FPGAs.
- roel_v 6y ago
  I'm not sure what part of 'the industry' you're referring to, but last time I looked (which was admittedly more than a year ago), there is only one serious SYCL implementation, it's proprietary and the only people pushing SYCL are... drum roll... the people selling that implementation. That's not even considering that OpenCL is in most cases (way) behind CUDA in performance, not to mention that the tooling for OpenCL is outright primitive. And all of that is not even mentioning that nvidia drags their feet massively in supporting newer OpenCL version on their GPU's. Maybe things have changed massively very recently, but I came back from iwocl (the OpenCL and SYCL conference) two years ago massively disappointed. And I still don't see anyone in the slice of the HPC space I'm familiar with (environmental & GIS-related modeling) using it, for a reason I presume.
  - rodburns 6y ago
    Intel are supporting SYCL in a big way. They are implementing SYCL in the open source LLVM project, the project is called DPC++. Argonne National Labs are also using SYCL to program the exascale supercomputer (Aurora) they are building. Whilst it's not the best solution for everyone there is certainly industry support for it. https://software.intel.com/content/www/us/en/develop/tools/oneapi/components/dpc-compiler.html https://software.intel.com/content/www/us/en/develop/tools/o... https://github.com/intel/llvm/blob/sycl/sycl/doc/GetStartedGuide.md https://github.com/intel/llvm/blob/sycl/sycl/doc/GetStartedG... https://www.alcf.anl.gov/support-center/training-assets/roadmap-sycldpc-aurora https://www.alcf.anl.gov/support-center/training-assets/road...
- paulsmal 6y ago
  this is nice, do you have any more details?
- freeone3000 6y ago
  Oh, the third option, with absolutely no library support. Or, looking at Khronos's page, even device support. At least OpenCL has the benefit of wide compatibility, even if its performance meant it was unsuitable. Nvidia has settled on CUDA, and has maintained 13 years of compatibility. If SYCL is going to ever compete with CUDA, it's going to need to stick around long enough for people to build on it.
  - throwaway56034 6y ago
    Well, Nvidia invented CUDA, it's no surprise that they won't give up their competitive advantage by properly supporting OpenCL or SYCL. Edit: Also, SYCL should be supported by most things already supporting OpenCL, like Tensorflow for example.
  - fortran77 6y ago
    My firm does GPGPU programming for audio processing. Nobody in this area even thinks of using anything other than CUDA. It has far more support, tools, compilers, etc.
    - Aaargh20318 6y ago
      What about the Mac side of things ? Mac's are still pretty big in audio production, but they don't support CUDA (and don't come with Nvidia GPUs any more anyway). How does Metal Compute stack up ?
      - fortran77 6y ago
        We don't use Macs. While many uusicians use macs, the people on the engineering/scientific side of things don't. We're not doing music audio.
- dragontamer 6y ago
  I think OpenMP device offload has a better chance than SYCL or OpenCL. But I'm not in the industry. CUDA's single-source environment is very good compared to SYCL or OpenCL. OpenMP still has single-source (just #pragma target), and is already adopted by the HPC crowd for multithreaded programming. NVidia and AMD support seems to be growing, especially because OpenMP is important to the supercomputer folks. (Summit and Frontier) ROCm is CUDA-like and mostly works. The main issue is that CUDA will always remain a few steps ahead as AMD is forced to play "catchup". (ex: Cooperative groups are useful but unimplemented for now in ROCm). ROCm's main issue is being stuck in Linux, they need some Windows support if they want ROCm to really take off.
- shmerl 6y ago
  Didn't they plan to merge OpenCL and Vulkan?
arnon 6y ago
ROCm and HIP - https://github.com/ROCm-Developer-Tools/HIP https://github.com/ROCm-Developer-Tools/HIP, as it has been for the last 3-4 years
- rrss 6y ago
  unless you have navi, in which case rocm is not supported and doesn't work and I guess AMD expects you to use opencl.
  - dragontamer 6y ago
    Navi just began to compile a week or two ago. Still 'unsupported', but its slowly gaining features. ROCm is a bit uneven with what cards it supports. Ex: Rx 550 never was supported (even though Polaris, the rest of the 5xx series worked). It seems like the only cards ROCm works for are the ones that share a chip and/or driver with AMD's "Machine Intelligence" line of cards. (MI50, MI60, etc. etc.). Which are Fiji (Rx Fury), Polaris (Rx 580), Vega, and Radeon VII. CUDA is a "works on all NVidia devices" kind of thing. So I think people are surprised when they have to read the docs for AMD's ROCm.
    - numpad0 6y ago
      Last time I tried, PCIe 3.0 was also a requirement, although that wasn't much of an issue unless it was some nonstandard configuration like an RX 580 in a junkyard find A4 APU machine.
    - Vvector 6y ago
      > Navi just began to compile a week or two ago. Still 'unsupported', but its slowly gaining features. 14 months after release, still unsupported. AMD needs to step up to the plate if they want to be seriously considered in the AI world.
      - dragontamer 6y ago
        Its pretty clear what's going on: AMD doesn't have the resources to support all the GPUs they release on ROCm. They focus on the "MI" line of cards, which have similarities to the consumer GPUs. (Ex: Rx Fury is very similar to MI8). Because MI8 and Rx Fury have similarities, AMD gives ROCm support to both. Ditto with MI25 and Vega, and MI6 and Rx 580. EDIT: With "Navi" gaining features in ROCm repos, my bet is that a new MI-card based on RDNA2 might be coming out soon. Maybe its Arcturus (the Exascale GPU for the Frontier supercomputer).

reitzensteinm 6y ago

Chips are in the pipeline for many years, so products being released today are probably on the tail end of being impacted by AMD cutting GPU budgets to the bone to focus on Zen. I wouldn't expect miracles. Given AMD's market cap and financial success, I bet they're working on competitive designs now. There's no money in being an also ran in the space.

DisjointedHunt 6y ago
The information available around Big Navi (The GPUs expected to launch next month) put them at a +50% performance jump from the last generation and likely able to compete with the Nvidia flagship 3080: https://www.tomshardware.com/news/amd-big_navi-rdna2-all-we-know https://www.tomshardware.com/news/amd-big_navi-rdna2-all-we-... I'd really urge anyone wanting to opine on these issues to watch the official analyst day videos and read through the releases AMD has put out so far in terms of architecture and other tech decisions they've made: https://www.youtube.com/watch?v=LMNwLVJQzGs&t=10064s https://www.youtube.com/watch?v=LMNwLVJQzGs&t=10064s
- mbell 6y ago
  > The information available around Big Navi (The GPUs expected to launch next month) put them at a +50% performance jump from the last generation and likely able to compete with the Nvidia flagship 3080 I guess it depends on what benchmark you care about but for AMD to compete with a 3080, which is 70-100% faster than a 2080, which is 10-20% faster than AMD's current flagship, they need to do a lot better than +50%.
  - strictnein 6y ago
    Some of the leaked benchmarks for the 3080 put it more in the range of 30-50% better than the 2080 series: "30% Faster Than The RTX 2080 Ti & 50% Faster Than the RTX 2080 SUPER on Average" https://wccftech.com/nvidia-geforce-rtx-3080-graphics-card-benchmarks-game-performance-leak/ https://wccftech.com/nvidia-geforce-rtx-3080-graphics-card-b...

deeeeeplearning 6y ago

>gaining market dominance, both in the ML/GPGPU space Is this really the case? Last I checked all the major Deep Learning toolsets ran off CUDA/cuDNN and there was nothing comparable for AMD hardware.

travisporter 6y ago
I agree. I think OP was referring to nvidia?
d0mdo0ss 6y ago
anyone have some experience on this? I'm in the market for an AMD laptop capable of running standard sci-kit/pytorch/etc but these seem optimized for NVIDIA cards. I'm curious about the outlook for these trending AMD cards.
- jandrese 6y ago
  Generally if you want to do ML tasks you want nVidia. They put the work in early to build the tooling so now the default assumption is that you're on nVidia hardware. It is possible to do some stuff on AMD cards, but you'll be on the cutting edge for that platform re-solving problems that were already solved on the nVidia side.
  - d0mdo0ss 6y ago
    yes that was my conclusion after some research on this. there is an open issue on github for AMD support on pytorch, and looks like something works on arch linux, but really sounds like support is still in the hacking stage and far from production mode. https://github.com/pytorch/pytorch/issues/10657 https://github.com/pytorch/pytorch/issues/10657
    - t-vi 6y ago
      Support for PyTorch on ROCm is fairly good. I have built it from the dev branch without much trouble for all year. There has been ROCm CI even longer, my patch to print ROCm system information for bug reports was merged last week. If you know where to look (i.e. it's public but unannounced) you can see that nightly wheels have been built for the last few days. So I would expect that some time between now and the Developer Day in November we'll see ROCm appear on PyTorch's "get started" page.

marmaduke 6y ago

I recently ported a code to AVX512 (with ISPC) and was blown away by the performance. Sure, maybe fma can cause throttling but for me, I could replace a 28 core Haswell machine by a 10 core Skylake X (Xeon W specifically) with the new instruction set. A fatter one (16 core?) would compete with a high end Nvidia card and maybe a 32 core EPYC. Sure, it’s just my benchmark, but it is competitive.