Announcement

Collapse
No announcement yet.

AMD "GFX908" Additions Land In LLVM 9.0 For New Workstation GPU

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

  • AMD "GFX908" Additions Land In LLVM 9.0 For New Workstation GPU

    Phoronix: AMD "GFX908" Additions Land In LLVM 9.0 For New Workstation GPU

    Weeks ahead of SIGGRAPH and days ahead of the LLVM 9.0 code branching, a number of big "GFX908" commits have been landing in the AMDGPU LLVM shader compiler back-end over the past day...

    http://www.phoronix.com/scan.php?pag...-LLVM-Activity

  • #2
    Well. there s Idea


    from article of June, the 22nd, 2017
    https://diit.cz/clanek/prezentace-k-vydani-epyc
    It is impossible to use Navi here
    Radeon ROCm 2.6 Released - Without Navi Support But Adds BFloat16 & Other Features
    9 July 2019
    https://www.phoronix.com/scan.php?pa...m-2.6-Released

    So new Vega may be this compute GPU for Epyc socket, yet





    Comment


    • #3
      Originally posted by Peter Fodrek View Post
      So new Vega may be this compute GPU for Epyc socket, yet
      If there is a GPU that could run in a CPU socket, it'd have to be one with HBM2. Because, any decently-powerful GPU would be bandwidth-starved with even 8 channels of DDR4.

      This is basically what Intel did with the Knights Landing generation of Xeon Phi. They put 16 GB of MCDRAM / HMC memory in-package, to augment the 6-channels of DDR4 provided by the current Xeon server socket. Except, those had no UPI links, which prevented them from being paired in any kind of multi-socket configuration.
      Last edited by coder; 07-12-2019, 01:13 AM.

      Comment


      • #4
        Originally posted by coder View Post
        If there is a GPU that could run in a CPU socket, it'd have to be one with HBM2. Because, any decently-powerful GPU would be bandwidth-starved with even 8 channels of DDR4.
        Yes if there will be no HBM inside the GPU package.

        And additionally
        Confirming that it's a workstation product is references to ECC memory,
        may but may not as refer to 2nd generation Vega known as Vega 2x or 3rd generation Vega that may be known as Vega 3x


        As of Polaris
        AMD Radeon 400 series
        Release date June 29, 2016; 3 years ago
        These cards were the first to feature the Polaris GPUs,
        https://en.wikipedia.org/wiki/AMD_Radeon_400_series
        and
        AMD Radeon 500 series

        The Radeon RX 500 series is a series of graphics cards made by AMD. These cards are based on the fourth iteration of the Graphics Core Next architecture, featuring GPUs based on Polaris 30, Polaris 20, Polaris 11, and Polaris 12 chips.
        Release date April 18, 2017; 2 years ago (Polaris 20)
        https://en.wikipedia.org/wiki/AMD_Radeon_500_series
        AMD Polaris 30
        Released Nov 15th, 2018
        https://www.techpowerup.com/gpu-spec...olaris-30.g877

        So there is time for Polaris 40 and Vega 30 this year. And for next year for Navi 20. Vega 40, Polaris 50 and NextGen 10 GPUs

        Comment


        • #5
          https://github.com/llvm-mirror/llvm/...6992207b4d38fa Wide loads
          https://github.com/llvm-mirror/llvm/...6992207b4d38fa MI-100

          Comment


          • #6
            My guess it that it's a variant with multiple GPU dies. Much like what is going to be used in the new Mac Pro (not released yet). Might even be the same card for the most part. MI-100 would suggest that it's part of the MI family of GPUs (MI50 and MI60 are both Vega 20 aka gfx906).
            Maybe gfx908 is 906 but with some extra refinements,

            Comment


            • #7
              MI-100 - it's a dual-GPU Vega 20, much like the dual-GPU Vega 20 in the future Mac Pro.

              Comment

              Working...
              X