Announcement

Collapse
No announcement yet.

OpenCL 3.0 Specification Released With New Khronos Open-Source OpenCL SDK

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

  • OpenCL 3.0 Specification Released With New Khronos Open-Source OpenCL SDK

    Phoronix: OpenCL 3.0 Specification Released With New Khronos Open-Source OpenCL SDK

    Back in April was the provisional release of OpenCL 3.0 with making CL 2.x features optional while adding async DMA extensions and more. Today the finalized version of OpenCL 3.0 has been released plus also introducing an official Khronos OpenCL SDK...

    http://www.phoronix.com/scan.php?pag...0-Released-SDK

  • #2
    i.e. rebranded opencl 1.2 with async dma extensions was released. wow huge progress

    Comment


    • #3
      If they make some more features optional, maybe OpenCL 4.0 will run on my old Riva TNT2 card?

      Comment


      • #4
        Originally posted by Mathias View Post
        If they make some more features optional, maybe OpenCL 4.0 will run on my old Riva TNT2 card?
        Why bother? I'm sure a modern AM4 or LGA 1200 CPU could perform compute calculations faster than the TNT2.

        Comment


        • #5
          What does this mean for HSA?

          Comment


          • #6
            Why nVidia would want to prepare OpenCL 3.0 driver when they have monopoly with proprietary CUDA? Maybe as OpenCL->CUDA (not very performance efficient) wrapper...

            Comment


            • #7
              Originally posted by Mathias View Post
              If they make some more features optional, maybe OpenCL 4.0 will run on my old Riva TNT2 card?
              I asked the same question back in April here: https://github.com/KhronosGroup/Vulkan-Docs/issues/1259

              Comment


              • #8
                Originally posted by schmidtbag View Post
                Why bother? I'm sure a modern AM4 or LGA 1200 CPU could perform compute calculations faster than the TNT2.
                He was just joking but some applications (resolve I'm looking at you) won't run openCL if it is on the CPU.

                With AMD (and NVidia) having upped their compute core counts considerably on their new graphics cards I am hoping the distros will get serious about pulling in the changes AMD has made to get ROCm usable for the every day man. It is starting to look like FC97 before Fedora will pull the changes in.

                Comment


                • #9
                  Originally posted by Mathias View Post
                  If they make some more features optional, maybe OpenCL 4.0 will run on my old Riva TNT2 card?
                  It's a valid feature given that OpenCL 3.0 is so powerful it can send you back in time and there you'll have to use old hw.
                  In fact it's so powerful that everyone using it is in the past and to us mere mortals it looks dead for years already.

                  Comment


                  • #10
                    Originally posted by wswartzendruber View Post
                    What does this mean for HSA?
                    HSA is more or less dead. It's been deprecated for ROCm by AMD.

                    HSA was more or less a cheerleading effort by AMD along with ARM and family members to spearhead development of Heterogeneous computing across different silicon assets such as CPUs, GPUs, DSPs, FPGAs, NPUs and now DPUs, and the frameworks and tools to start doing that. Obviously one big part of that was OpenCL. However, at the time years ago when HSA was founded, only AMD was really pushing OpenCL, along with Apple but for their ecosystem that had nothing really to do with HSA. Nvidia had and still does push CUDA. Intel had no GPU assets and was still pushing the HPC world into their "Many Cores" paradigm with Knights Landing and pushing for POCL because...well....no Intel GPU.

                    NOW.....things have REALLY changed. Intel poached an AMD head Radeon engineer and is going full guns blazing into GPUs. They are tooling up support for OpenCL proper and not just a portable software CPU based version of it. EVERYONE has their own version of some heterogeneous computing scheme and even the Linux kernel is getting wired up with HMM (Heterogeneous Memory Management) which will provide a more generic version of what HSA as spearheaded by AMD was trying to accomplish. Not to mention the CXL and GenZ protocols as extensions of PCIe that move bits and bobs around internally on the mainboard and externally outside to the rack, HSA has simply been bypassed.

                    You could look at HSA as the Heterogeneous Architecture version of AMD's "3DNow" initiative back in the late 90's to one up Intel's MMX SIMD protocol. It had some nice things going for it. And a few games actually took advantage of it for some benefit above and beyond MMX. But the industry as a whole never accepted it and they eventually ignored it and moved on as Intel expanded MMX into SSE and later AVX. Now with CXL, GenZ, proper support for OpenCL from intel with actual Intel GPUs, HMM in the Linux Kernel, etc....there's not much need for an AMD spearheaded initiative in the form of HSA seeing as how, as I said above, even AMD has moved on with ROCm.

                    Also it doesn't help that AMD still makes ROCm a shit show to install and setup. Not to mention that they have all but abandoned HSA / ROCm support on the very APU platform that they started the whole movement with....the Bullldozer/Steamroller/Excavator cores. Kaveri is really the only one. Not so much with Carrizo and Bristol Ridge which are superior to Kaveri. HSA seems now, in hindsight, to be a bit of "bait and switch". Like AMD was pulling an "Intel" with their marketing of HSA. All that "FUSION" and HSA bullshit....remember that ??

                    Comment

                    Working...
                    X