Announcement

Collapse
No announcement yet.

Intel CR 24.17.29377.6 Offers Latest OpenCL & oneAPI Level Zero Support

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

  • Intel CR 24.17.29377.6 Offers Latest OpenCL & oneAPI Level Zero Support

    Phoronix: Intel CR 24.17.29377.6 Offers Latest OpenCL & oneAPI Level Zero Support

    Intel Compute Runtime 24.17.29377.6 is now available as the latest routine update to this open-source GPU compute stack used by the company's integrated and discrete graphics products for providing OpenCL and oneAPI Level Zero compute capabilities...

    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
    I am glad they still support Skylake. AMD, where are you?

    Comment


    • #3
      Originally posted by tildearrow View Post
      I am glad they still support Skylake.
      Intel offers a long-life qualification, on some of their CPU models. I expect that has a lot to do with it.

      Comment


      • #4
        I'm watching with trepidation whether there'll be some bugs / performance regressions on DG2 et. al. due to the recent i915 "stable" kernel propagated changes which have AFAICT broken compute on DG2 and apparently subsequently caused some performance regression.
        Hopefully the level-zero / opencl / syscl updates needed if any were synchronized with the latest upcoming fixes for kernel 6.9 / 6.6 stable etc.

        Comment


        • #5
          Originally posted by coder View Post
          Intel offers a long-life qualification, on some of their CPU models. I expect that has a lot to do with it.
          Likely, perhaps, that the graphics architecture range (Gen 9 - Gen 9.5) is the same that was used from the Intel HD 500 in Apollo-Lake (Skylake-era) Atom CPUs through the Core i5-10600 Coffee-Lake Refresh-era desktop CPU. Yes only specific ones are listed as officially supported but yes; most seem to work anyway in various devices and configurations outside the supporedt zone; so that's cool.
          --> If extended processor lifecycles ​were the consideration; these projects would have included a larger number of CPUs with iGPUs listed; such as the Gen 7 Ivy Bridge, Gen 7.5 in Haswell Gen 8 in Broadwell. Gen8 even has FP16 support.<--

          • Ivy Bridge: Introduced 12 years and 3 weeks ago: (29 April 2012)
            • Final Line EOL: (30 Jun 2020)
          • ​Haswell: Introduced 10 years and 11 months ago: (4 June 2013)
            • Final Line EOL: (30 June 2021)
          • Broadwell: Introduced 9 years and 6 months ago (27 October 2014)
            • Final Line EOL: (30 June 2022)​


          So: Gen7-8 were relegated to software-unsupported status quite a long time ago, long before EOL, and have to use the beignet driver stack on *nix for OpenCL 2.1; even the Mesa "rusticl" driver doesn't cover them as they are part of the "clover" driver stack.

          (Side note: the Vulkan drivers were never finished either).

          (SideSide Note: I truly believed for a time that OneAPI would unify The Gen 7 through modern GPU compute stack with as many capabilities as possible for each). So, OpenCL can still be used for whatever workloads, including LLMs, but the driver hasn't been updated in quite some time and needs an older version of LLVM to build unless someone writes a patch. Business desktops with IB/HSW/BDW CPUs and 32GB RAM make for very cheap local LLM tinker machines... which would only be enhanced with a modern driver stack that is in a great part proprietary. Lightweight games and emulation could be had with a completed Vulkan driver. I think part of the issue here is no one realized how popular Vulkan and OpenCL and compute in general would be. With the release of NPUs, I worry that GPUs overall will be treated as "second-class citizens" of the computing world in terns of drivers: just like the original, innovative CPUs/iGPUs in the "modern Intel Core" lineup have seen once the success of Skylake really took off.)

          Comment


          • #6
            Originally posted by Eirikr1848 View Post
            Likely, perhaps, that the graphics architecture range (Gen 9 - Gen 9.5) is the same that was used from the Intel HD 500 in Apollo-Lake (Skylake-era) Atom CPUs through the Core i5-10600 Coffee-Lake Refresh-era desktop CPU. Yes only specific ones are listed as officially supported but yes; most seem to work anyway in various devices and configurations outside the supporedt zone; so that's cool.
            --> If extended processor lifecycles ​were the consideration; these projects would have included a larger number of CPUs with iGPUs listed; such as the Gen 7 Ivy Bridge, Gen 7.5 in Haswell Gen 8 in Broadwell. Gen8 even has FP16 support.<--

            • Ivy Bridge: Introduced 12 years and 3 weeks ago: (29 April 2012)
              • Final Line EOL: (30 Jun 2020)
            • ​Haswell: Introduced 10 years and 11 months ago: (4 June 2013)
              • Final Line EOL: (30 June 2021)
            • Broadwell: Introduced 9 years and 6 months ago (27 October 2014)
              • Final Line EOL: (30 June 2022)​


            So: Gen7-8 were relegated to software-unsupported status quite a long time ago, long before EOL, and have to use the beignet driver stack on *nix for OpenCL 2.1; even the Mesa "rusticl" driver doesn't cover them as they are part of the "clover" driver stack.

            (Side note: the Vulkan drivers were never finished either).

            (SideSide Note: I truly believed for a time that OneAPI would unify The Gen 7 through modern GPU compute stack with as many capabilities as possible for each). So, OpenCL can still be used for whatever workloads, including LLMs, but the driver hasn't been updated in quite some time and needs an older version of LLVM to build unless someone writes a patch. Business desktops with IB/HSW/BDW CPUs and 32GB RAM make for very cheap local LLM tinker machines... which would only be enhanced with a modern driver stack that is in a great part proprietary. Lightweight games and emulation could be had with a completed Vulkan driver. I think part of the issue here is no one realized how popular Vulkan and OpenCL and compute in general would be. With the release of NPUs, I worry that GPUs overall will be treated as "second-class citizens" of the computing world in terns of drivers: just like the original, innovative CPUs/iGPUs in the "modern Intel Core" lineup have seen once the success of Skylake really took off.)
            Wow, thanks for all the info! I agree with your takes, as well.

            Comment

            Working...
            X