Linux 6.14 To Bring An Important Improvement For AMD Preferred Core

Collapse
X
 
  • Filter
  • Time
  • Show
Clear All
new posts
  • phoronix
    Administrator
    • Jan 2007
    • 67385

    Linux 6.14 To Bring An Important Improvement For AMD Preferred Core

    Phoronix: Linux 6.14 To Bring An Important Improvement For AMD Preferred Core

    Being merged back in the Linux 6.9 kernel was AMD Preferred Core support within the amd_pstate driver for being able to communicate the "preferred" cores to the kernel for those that are able to reach a higher maximum frequency or otherwise be preferred over other CPU cores. For the upcoming Linux 6.14 merge window, an important set of patches are queued up for better positioning this Preferred Core handling...

    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
  • Kjell
    Senior Member
    • Apr 2019
    • 691

    #2
    CCX aware load balancing to retain warm L3 cache would help clear up CPU jitter, there's papers on this but I haven't seen much interest upstream

    Comment

    • Kjell
      Senior Member
      • Apr 2019
      • 691

      #3
      Originally posted by Kjell View Post
      CCX aware load balancing to retain warm L3 cache would help clear up CPU jitter, there's papers on this but I haven't seen much interest upstream
      For reference, this would also improve draw call performance in graphic intensive workloads/gaming

      Comment

      • Noitatsidem
        Phoronix Member
        • Sep 2015
        • 118

        #4
        Originally posted by Kjell View Post

        For reference, this would also improve draw call performance in graphic intensive workloads/gaming
        by how much?

        Comment

        • ptr1337
          Senior Member
          • Dec 2021
          • 216

          #5
          Very nice to see this finally merged!
          This will fix the preffered core feature and the 3D Cache driver will also work dynamically. Great work!

          Comment

          • Errinwright
            Senior Member
            • Aug 2023
            • 189

            #6
            Originally posted by Kjell View Post

            For reference, this would also improve draw call performance in graphic intensive workloads/gaming
            Perhaps the scx_rustland devs can integrate this somehow...?

            Comment

            • ptr1337
              Senior Member
              • Dec 2021
              • 216

              #7
              Originally posted by Errinwright View Post

              Perhaps the scx_rustland devs can integrate this somehow...?
              There are already scx scheduler, which are for this:
              1. scx_flash
              2. scx_bpfland (got more converted to a throughput scheduler for cuda and co, but still is quite good on desktops)
              3. scx_lavd

              Comment

              • F.Ultra
                Senior Member
                • Feb 2010
                • 2052

                #8
                Originally posted by Errinwright View Post

                Perhaps the scx_rustland devs can integrate this somehow...?
                a problem here is that the info that a core is part of a specific ccx is not exported anywhere.

                Comment

                • Kjell
                  Senior Member
                  • Apr 2019
                  • 691

                  #9
                  Originally posted by Noitatsidem View Post
                  by how much?
                  The idea is to reduce latency and subsequently eliminate micro-stuttering. The biproduct can be improved FPS which would be substantial in a CPU bound situation, around a +1.7% - +5.4% FPS uplift.

                  This can be partly achieve through CPU Affinity (here's some benchmarks of this concept: https://youtu.be/5LVsh61r3E0). However, the proper implementation is a lot more nuanced:

                  You need to factor in CPU's topology, the CCD+CCX layout to keep L3 caches warm and ideally by using core pairs to benefit from L2 cache.

                  To get a better idea why this matters, check Ryzen latency chart here:



                  This is why I disable CPPC Preferred Cores in BIOS, it introduces spikes in frametimes.

                  Comment

                  • nuetzel
                    Senior Member
                    • May 2016
                    • 754

                    #10
                    +1
                    We need upvote. Michael

                    Comment

                    Working...
                    X