Announcement

Collapse
No announcement yet.

Radeon ROCm 3.8 Released With Hipfort For Fortran On GPUs, Data Center Tool

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

  • #11
    For the newer APUs we switched from using APU paths (memory accessed via ATC and IOMMUv2) to using dGPU paths (memory accessed via GPUVM).

    In the short term you lose the ability to access unpinned memory (ie the ability to access OS-allocated memory without an initial API call) but it should make APUs a better development and testing vehicle for dGPU deployment and should make users less dependent on vendors implementing the correct SBIOS logic (eg CRAT table).

    I believe this started with Picasso aka 3000 series.
    Last edited by bridgman; 22 September 2020, 11:38 PM.
    Test signature

    Comment


    • #12
      Is there any support for calling xGEMM LAPACK routines using the Fortran interface? That would make it awesome.

      Comment


      • #13
        So, in summary, latest AMD GPUs work unreliably, APUs work unreliably, ROCm making progress in software but not in hardware support?

        Will RDNA2 be supported (and stable)? Or should I just resign myself to being tied to nVidia forevermore?

        Comment


        • #14
          I would tweak your summary to "supported AMD GPUs work reliably, not-yet-supported AMD GPUs work unreliably, new APUs will be running same code paths as dGPUs so they will leverage all the dGPU work, and HIP support for APUs is still being worked on".

          We are working on RDNA support as a lower priority project; trying to secure staffing to make it a high priority so it can happen quickly. For the last few years the ROCm focus has been on datacenter parts but that is gradually changing.
          Last edited by bridgman; 24 September 2020, 11:51 PM.
          Test signature

          Comment


          • #15
            I got my R470 work with tf for simple networks but as soon as it gets a little more deep strange things happens and learning doesn't start (issue reported for MiOpen, but not heard back). But if it works out seems quite fast. Oh, and works with upstream kernel from Arch, which is nice. So I'm positive they will eventually get there, especially since they need it working for CDNA2.
            An RX 470 uses the gfx803 ISA, which has loads of issues under ROCm. I collected them here:

            https://github.com/ROCmSoftwarePlatf...eam/issues/479

            I have Vegas (gfx900) and they work flawlessly with TensorFlow, so until the support of the 5000 (gfx10) series matures, these seem to be the way to go.

            Comment


            • #16
              Originally posted by bridgman View Post
              I would tweak your summary to "supported AMD GPUs work reliably, not-yet-supported AMD GPUs work unreliably, new APUs will be running same code paths as dGPUs so they will leverage all the dGPU work, and HIP support for APUs is still being worked on".

              We are working on RDNA support as a lower priority project; trying to secure staffing to make it a high priority so it can happen quickly. For the last few years the ROCm focus has been on datacenter parts but that is gradually changing.
              Thanks. I'm just grumpy 'cause nVidia drivers have been winding me up recently.

              Comment

              Working...
              X