Announcement

Collapse
No announcement yet.

Radeon ROCm 1.8.2 Compute Stack In Beta, Might Work Under Ubuntu 18.04 LTS

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

  • Radeon ROCm 1.8.2 Compute Stack In Beta, Might Work Under Ubuntu 18.04 LTS

    Phoronix: Radeon ROCm 1.8.2 Compute Stack In Beta, Might Work Under Ubuntu 18.04 LTS

    A new beta of the Radeon Open Compute "ROCm" stack was quietly made available for v1.8.2...

    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
    it looks like the ROCm 1.8.2 beta might contain preliminary Ubuntu 18.04 LTS "Bionic Beaver" support. A ROCm 1.8.2 beta user has commented that he was able to get 1.8.2 working on Ubuntu 18.04 with the Linux 4.16 kernel with the AMDKFD kernel driver.
    I have it working Im posting instructions and howto but the script's online its a custom patched kernel. With the latest patches merged back to 4.16 from freedesktop.orgs git. I used the tarball then disassembled parts and replace the kernel dkms module with a shim.
    Looking at the comment, the user appears to have not used the kernel drivers from ROCm 1.8.2 - and it's the kernel drivers that mostly drive OS/kernel compatibility. They didn't mention which hardware they are running, but they are running a custom kernel - 4.16 with latest KFD support backported from upstream. They have published instructions etc... which is nice, but what they have is a custom upstream-based kernel NOT Ubuntu 18.04 support in ROCm 1.8.2.

    POR for 18.04 support is still ROCM 1.9 AFAIK.
    Test signature

    Comment


    • #3
      I'm glad AMD is improving their machine learning/compute stack, but at the same time it's really sad 99% of machine learning solutions out there have essentially zero support for AMD products and software.

      Im working on something that needs machine learning and I will probably end up training it on my Ryzen rather than on my Polaris GPU.

      Comment


      • #4
        Would be great if ROCm would add extension to support the amd instructions to permute data between work items in a wavefront in OpenCL. Intel has one, but you have to use asm for Amd.

        Comment


        • #5
          Originally posted by Aeder View Post
          Im working on something that needs machine learning and I will probably end up training it on my Ryzen rather than on my Polaris GPU.
          Tensorflow 1.3 works well on AMD GPU using ROCm. Latest Tensorflow and PyTorch should be out soon with 1.8.2.

          Comment


          • #6
            Originally posted by bsp2020 View Post
            Tensorflow 1.3 works well on AMD GPU using ROCm. Latest Tensorflow and PyTorch should be out soon with 1.8.2.
            Tensorflow 1.8 for ROCm is public now:

            Test signature

            Comment

            Working...
            X