Announcement

Collapse
No announcement yet.

ZLUDA: Drop-In Open-Source CUDA Support For Intel Xe / UHD Graphics

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

  • ZLUDA: Drop-In Open-Source CUDA Support For Intel Xe / UHD Graphics

    Phoronix: ZLUDA: Drop-In Open-Source CUDA Support For Intel Xe / UHD Graphics

    An interesting solution built off Intel's oneAPI Level Zero is the open-source "ZLUDA" that is providing a "Level Zero CUDA" implementation for being able to run programs geared for NVIDIA CUDA atop Intel UHD / Xe Graphics hardware...

    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
    As people seem to love programming for CUDA and the fragmentation between the 3 GPU producers in terms of API gets bigger and bigger. Maybe CUDA running everywhere is not a bad idea.

    Comment


    • #3
      Will it work with AMD too? Breaking CUDA lock-in will be a huge boon.

      Comment


      • #4
        I am worried that NVIDIA may end up suing Intel...

        Comment


        • #5
          Well if can be supported legaly by the hw vendor on all platfoms. If not cuda copatibility is bad.

          SW vendors should use khoronos standards instead, other than that they jail themselve. Though, this might be ok for shortsighted managements.

          Comment


          • #6
            Originally posted by Alexmitter View Post
            As people seem to love programming for CUDA and the fragmentation between the 3 GPU producers in terms of API gets bigger and bigger. Maybe CUDA running everywhere is not a bad idea.
            Problem here would be that Nvidia dictates every single inch of future of CUDA. That being said, there is no good mature alternative, except maybe Vulkan compute in future.

            Comment


            • #7
              Originally posted by tildearrow View Post
              I am worried that NVIDIA may end up suing Intel...
              For what? Some user developing his own CUDA to zero API converter and running that on a Intel GPU? LOL, have you even read the article

              Comment


              • #8
                Originally posted by shmerl View Post
                Will it work with AMD too? Breaking CUDA lock-in will be a huge boon.
                Well, you can convert CUDA source code via HIP to make it run on ROCm (or CUDA). That is working for ages. It is, however, not a drop-in replacement, as things need to be recompiled. Without recompilation, I fear, it would hurt performance a lot. That's probably why it wasn't done already.

                Comment


                • #9
                  Originally posted by Alexmitter View Post

                  For what? Some user developing his own CUDA to zero API converter and running that on a Intel GPU? LOL, have you even read the article
                  I know, right, that's about as absurd as the USA suing a New Zealand online storage hosting provider over what some user uploaded there.

                  Oh wait...

                  One of the main reasons to opt for cuda over opencl is the amount of libraries nvidia themselves provide as a convenience to promote their own proprietary thing. Somehow I doubt they will be thrilled about letting other hardware vendors into that pot.
                  Last edited by ddriver; 24 November 2020, 06:34 AM.

                  Comment


                  • #10
                    Originally posted by Alexmitter View Post
                    As people seem to love programming for CUDA and the fragmentation between the 3 GPU producers in terms of API gets bigger and bigger. Maybe CUDA running everywhere is not a bad idea.
                    CUDA is actually the easiest GPU programming API to work with.
                    Most alternatives (OpenCL/SyCL/oneAPI/etc) are all fundamentally broken. Some of them are verbose, some of them are not intuitive.

                    The only API that looks Okay to me is HIP, but basically this is still CUDA.
                    Last edited by zxy_thf; 24 November 2020, 07:38 AM.

                    Comment

                    Working...
                    X