Announcement

Collapse
No announcement yet.

AMD Revises Work On GPU Workload Hinting For Linux

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

  • AMD Revises Work On GPU Workload Hinting For Linux

    Phoronix: AMD Revises Work On GPU Workload Hinting For Linux

    AMD's Linux graphics driver engineers this week sent out their latest patch series working on workload hints that are used for dynamically tuning the power profile of AMD GPU SoCs based upon the specified workload indicator...

    https://www.phoronix.com/news/AMDGPU-Workload-Hints-v2

  • #2
    Seems pretty obvious, in hindsight. I do wonder about some of the assumptions behind it, though. Presumably, VIDEO is trying to run the GPU just fast enough to keep from running behind, while VR is juicing the boost clocks for minimum latency. Perhaps COMPUTE converse limits the amount of boosting for better efficiency.

    Comment


    • #3
      And what if somebody runs a compute job in the background and watches a movie in the foreground?
      Or plays a game on the main screen and runs a movie on the attached TV screen?
      The second example I already did a few times and everything worked great, but I was running Windows at that time.

      Comment


      • #4
        Originally posted by Danny3 View Post
        And what if somebody runs a compute job in the background and watches a movie in the foreground?
        Or plays a game on the main screen and runs a movie on the attached TV screen?
        The second example I already did a few times and everything worked great, but I was running Windows at that time.
        Presumably this is a per application setting, and the driver will activate the parts of the chip needed for every running application. Also, It could simply recieve it at every job submission, so it changes dynamically depending on the job.

        Comment


        • #5
          Originally posted by dragorth View Post
          Presumably this is a per application setting,
          The patch seems to be configuring it on something called a "workload context". I'm not sure exactly what that corresponds to, in userspace APIs. I think it's reasonable to assume it's at least process-level granularity, if not finer.

          Comment


          • #6
            Originally posted by dragorth View Post

            Presumably this is a per application setting, and the driver will activate the parts of the chip needed for every running application. Also, It could simply recieve it at every job submission, so it changes dynamically depending on the job.
            At a quick glance, it looks to me like it's set on every job submission to the kernel and then gets reset as soon as the job completes. So that would be way more fine-grained than just per application.

            Comment


            • #7
              Wonder if its possible to have that windows enhanced profile under Linux that is meant to increase performance by %13 or something... forgot the name of it.

              Comment

              Working...
              X