Announcement

Collapse
No announcement yet.

RADV Lands A Few More Improvements To Reduce CPU Overhead

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

  • RADV Lands A Few More Improvements To Reduce CPU Overhead

    Phoronix: RADV Lands A Few More Improvements To Reduce CPU Overhead

    Samuel Pitoiset of Valve's Linux graphics team has landed a few patches into Mesa 23.0 for further reducing the CPU overhead of the Vulkan driver's draw path...

    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
    Wonderful!
    Thank you very much AMD and Valve!

    Comment


    • #3
      I'm curious to see how it translates into benchmarks.

      Comment


      • #4
        I am very glad I got myself an RX 6000-series GPU, the good stuff just keeps coming. Thanks Valve and AMD!

        Comment


        • #5
          This is very cool news and shows there's always more perf to be squeezed out with the right tools and approach. There's been a huge amount of beneficial Mesa work in the past 3+ months and a lot of it coming within the upcoming 23.0.0 release. Let hope it's released on time and makes it into Debian and other early 2023 distros.

          Comment


          • #6
            Originally posted by spykes View Post
            I'm curious to see how it translates into benchmarks.
            yeah didn't he recently do some benches that looked cpu constrained? 11% in any case is huge

            Comment


            • #7
              This is great but it's especially important for Deck users, considering the limited CPU performance.

              Comment


              • #8
                Originally posted by doomie View Post

                yeah didn't he recently do some benches that looked cpu constrained? 11% in any case is huge
                I think it is just 11% reduction in overhead. The performance increase should be much more subtle.

                Comment


                • #9
                  Originally posted by JellyBrain View Post

                  I think it is just 11% reduction in overhead. The performance increase should be much more subtle.
                  The merge request says "+~11% with ./vkoverhead -test 0 -duration 15". If we use the number from Mike Blumenkrantz's README draw result as an example, +11% of 30595 for test 0... that's around 34 million draws/second vs 30.6 million.

                  Comment


                  • #10
                    Just to provide some context to the numbers: I get 33400-33900k draws in test 0 on a Gentoo installation that's tailored towards my CPU with mesa 22.3.2. (5900X, RX 6800XT). Kernel 6.1.2 w/ 1000hz tick, BORE scheduler and voluntary preemption.

                    A generic Arch installation with a Core i7 [email protected] and RX 470 manages 20000k draws calls.

                    Comment

                    Working...
                    X