Announcement

Collapse
No announcement yet.

AMD Quietly Funded A Drop-In CUDA Implementation Built On ROCm: It's Now Open-Source

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

  • #31
    Originally posted by Serafean View Post
    Awsome job!
    Now take it and stick it into Mesa alongside rusticl Cuda on freedreno would be fun.
    (I know it can't be so easy...)
    CUDA and RUSTICL on all Gallium3D drivers!

    ... But for real though, this entire effort feels like something Collabora would hire the dev for

    Comment


    • #32
      Originally posted by avis View Post
      "Złuda roughly means "delusion" / "mirage" / "illusion" in Polish, given the author is called Andrzej Janik this may be a pun " - source.
      Yes and CUDA means 'miracles'.

      Comment


      • #33
        Originally posted by and.elf View Post
        Very interesting, indeed. How can a library wrapper be faster than a native implementation? Is the rocm implementation poorly optimized, somehow? Or, did I understand this altogether?
        Blender is well known to have NVidia fanboys for developers. They are on the record stating that they never have, and never will, buy an AMD GPU to even TEST AMD code submitted to the Blender project, let alone write any themselves. They also actively try to remove AMD codepaths from the codebase every few years. So the people most familiar with the code only ever works with and optimizes the entire architecture for CUDA.

        It's why I hate seeing Blender used as a benchmark. These results alone prove that the AMD implementation in Blender is nowhere near the capability of the cards if a translation layer gets better results, so any benchmark comparing AMD and NVidia GPUs on it in terms of performance is moot. NVidia cards are considered "better for productivity" purely because developers of productivity applications only cater to CUDA and not ROCm.

        Which isn't to say that NVidia isn't better than AMD in a lot of ways. Their raytracing performance is miles better, and they have way better "secondary" applications for their GPUs and tensor cores, like RTX Voice, NVENC, etc. Just mildly annoying that we've never really seen how fast AMD GPUs *can* go in a productivity benchmark, because nobody's actually bothered trying to optimize for it.

        Comment


        • #34
          Originally posted by Daktyl198 View Post

          Blender is well known to have NVidia fanboys for developers. They are on the record stating that they never have, and never will, buy an AMD GPU to even TEST AMD code submitted to the Blender project, let alone write any themselves. They also actively try to remove AMD codepaths from the codebase every few years. So the people most familiar with the code only ever works with and optimizes the entire architecture for CUDA.
          Citations needed.

          Originally posted by Daktyl198 View Post
          It's why I hate seeing Blender used as a benchmark. These results alone prove that the AMD implementation in Blender is nowhere near the capability of the cards if a translation layer gets better results, so any benchmark comparing AMD and NVidia GPUs on it in terms of performance is moot. NVidia cards are considered "better for productivity" purely because developers of productivity applications only cater to CUDA and not ROCm.
          But Blender is open source, extremely popular and widely used and I don't know anything similar or close to it. What else would Michael benchmark?

          Originally posted by Daktyl198 View Post
          Which isn't to say that NVidia isn't better than AMD in a lot of ways. Their raytracing performance is miles better, and they have way better "secondary" applications for their GPUs and tensor cores, like RTX Voice, NVENC, etc. Just mildly annoying that we've never really seen how fast AMD GPUs *can* go in a productivity benchmark, because nobody's actually bothered trying to optimize for it.
          A chicken and egg problem considering AMD's market share.

          Comment


          • #35
            Hmm I wonder if AMD dropped this because of other possibilities like the Conda plug-in being developed for openmm .
            The early test results were quite promising. You can find the results in an earlier thread containing a google docs link.
            Final HIP Platform implementation for AMD GPUs on ROCm by AJcodes · Pull Request Noob question on installing propriet... with fedora 7 · openmm/openmm · GitHub​

            Contribute to StreamHPC/openmm-hip development by creating an account on GitHub.
            Those who would give up Essential Liberty to purchase a little Temporary Safety,deserve neither Liberty nor Safety.
            Ben Franklin 1755

            Comment


            • #36
              Originally posted by Anux View Post
              Not bad, why would AMD stop that project, especially considering the results?
              Because when you run a business it's never a good idea to adopt a competitor's flagship product.

              If AMD officially advertised that you can run CUDA code on top of AMD products then a prudent consumer would ask why not just buy an NVIDIA card?

              Same reason why Intel doesn't support the project.

              Comment


              • #37
                Originally posted by cb88 View Post
                That's definitely some sidestepping of legal issues... AMD paid for the guy to work for them he happened to write/improve ZLUDA and it happened to get released as open source on contingency of his contract with AMD ending.... that's definitely intentional but it also means that ZLUDA can't be directly tied back to AMD as a product. On the other hand AMD didn't accidentally hire this guy to work on just this fora whole year... sounds like it works quite well too.

                Sounds like even though AMD didn't release it officially hey really did the guy a solid.
                It is just further proof that AMD is an incompetent company - one guy has to use a 'version' or 'hack' of CUDA to increase/improve rendering performance with AMD gpus - ONE guy - AND AMD as a company couldn't utilize creative ideas of their own or their own tech/engineering to improve their software/hardware - also, btw, for Blender - Nvidia uses OptiX to improve performance beyond CUDA - something AMD has not been able to do with HIP.

                Perhaps, this one guy can help them there, too?

                Comment


                • #38
                  Now this just need to make rocm/hip stack actually stable and usable

                  Comment


                  • #39
                    Just in case anyone didn't want to wade through the Final hip thread to look for the google doc link here it is as you can see the cuda code run through hip is far more performant than OpenCL code.
                    OpenCL vs HIP (OpenMM8.0.0) - Google Sheets
                    Last edited by DarkFoss; 12 February 2024, 04:12 PM. Reason: wrong link
                    Those who would give up Essential Liberty to purchase a little Temporary Safety,deserve neither Liberty nor Safety.
                    Ben Franklin 1755

                    Comment


                    • #40
                      Originally posted by stormcrow View Post
                      Oracle v. Google was decided in 2021. That case effectively makes APIs copyrightable in the US. While Intel began the project, that SCOTUS case alters the copyright field in the US. In order to use APIs, you either have to comply with the license, or be able to successfully argue fair use in a trial, an expensive and lengthy process without guarantees of success. People shouldn't make assumptions what AMD knows and doesn't know, or what Nvidia sees as a problem or doesn't. It's up to a property owner when they wish to file suit against alleged infringers. Just because they've been silent doesn't mean they will remain silent - nor do we know what's going on behind the scenes when no one talks. Successful estoppel defenses are rare. You can't assume a property holder is ok because "surely they know". You have to actively engage with the property holder in good faith, then and only then if they agree to allow something then renege on the agreement does one have a strong case for estoppel.

                      You should never take a tech reporter's assumptions about the law at face value, especially when they're making assumptions about who knew what and when. Reporters are nearly always inaccurate when it comes to legal matters. (especially in the tech press!)
                      You are correct, however there is a principle in law called Estoppel.

                      This principle has been used successfully to fight claims of copyright infringement.

                      The idea is if a party has allowed the public to use their IP without restriction, they can not suddenly sue one party for violating your rights.

                      With regards to APIs, it would be like Microsoft after decades of allowing programs, regardless of the license they use of whether it is compatible with MS' licensing terms, to use win32 hooks into the OS, to suddenly sue the developers of one game because that specific game somehow does not comply with some licensing restriction.

                      An Estoppel defense would surely be raised and it's likely it would work.

                      Of course, it's theoretical and no one knows how a judge or jury will rule.
                      Last edited by sophisticles; 12 February 2024, 09:17 PM.

                      Comment

                      Working...
                      X