Announcement

Collapse
No announcement yet.

Intel oneAPI DPC++ Compiler Merges Its Initial CUDA Backend

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

  • Intel oneAPI DPC++ Compiler Merges Its Initial CUDA Backend

    Phoronix: Intel oneAPI DPC++ Compiler Merges Its Initial CUDA Backend

    Intel's oneAPI crew just released version 2020-03 (though one would have thought it should be 2020-05) of their Data Parallel C++ (DPC++) compiler and with this release are several new features including the NVIDIA CUDA back-end...

    Phoronix, Linux Hardware Reviews, Linux hardware benchmarks, Linux server benchmarks, Linux benchmarking, Desktop Linux, Linux performance, Open Source graphics, Linux How To, Ubuntu benchmarks, Ubuntu hardware, Phoronix Test Suite

  • #2
    Fascinating. A few weeks ago Michael ran an article on New! Improved!! Now with even more OpenCL!!! 3.0 https://www.phoronix.com/scan.php?pa...-30-spec&num=1... and the comments included discussion on how Intel and AMD might (hypothetically) compete with CUDA. Obviously, any shop or lab heavily invested in CUDA is not going to turn away from Nvidia anytime Real Soon. But stacking OneAPI on top of CUDA at least gives their devs an option to tinker with, and perhaps some future vendor heterogeneity.

    Following the links, it seems OneAPI's principle GPU interface is OPenCL 2.1, for an appropriate definition of that standard, so Vega can also play in the game if/when AND has suitable drivers. https://github.com/intel/llvm/releases/tag/2020-03

    (I don't know if they do already or not, one way or the other. Ask Andy?)

    Comment


    • #3
      Originally posted by pipe13 View Post
      But stacking OneAPI on top of CUDA at least gives their devs an option to tinker with, and perhaps some future vendor heterogeneity.
      I think the issue with switching away from CUDA is losing access to the CUDA libraries like CUDNN and CUBLAS which are heavily optimised for NVIDIA GPUs


      Comment


      • #4
        I wonder which version of CUDA is required. I have a good Maxwell-generation card that I'd potentially be interested in using to kick the tires of oneAPI, but I'm not sure the latest CUDA even supports it.

        Comment


        • #5
          Originally posted by coder View Post
          I wonder which version of CUDA is required. I have a good Maxwell-generation card that I'd potentially be interested in using to kick the tires of oneAPI, but I'm not sure the latest CUDA even supports it.

          I'm quite sure CUDA 11 supports anything from CUDA compute capability 5.2 onwards, and 2nd Gen Maxwell supports this. So that's anything with a GM200, 204 OR 206 GPU.

          If I were to guess, I would think that CUDA 12 might also support Maxwell, but after that I'm not sure.

          Comment

          Working...
          X