Announcement

Collapse
No announcement yet.

RadeonSI/Gallium3D Still Appears To Have Greater CPU Overhead Than The NVIDIA Driver

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

  • #11
    I bet Marek is looking through his code now trying to figure out how to squeeze a few percent more. I'll be disappointed if Mesa isn't faster than the nvidia pro driver in the next few months :-P.

    Impressive work guys. I can't wait for Vega to hit the stores and the DC/DAL to accepted upstream. I've just bought a new freesync monitor and I'm looking forward to the time I can use it with Mesa.

    Comment


    • #12
      Originally posted by marek View Post
      OMG it's so close!
      Agreed.. I've been watching all your draw overhead patches, along with the threaded gallium changes, with interest.

      Between your, Nicolai, Timothy, and Samuel's work, there's been lots to be excited about in the radeonsi driver recently.

      Keep it up, it's definitely appreciated.

      Comment


      • #13
        Introducing 0.1% and 1% lows from time to time would be beneficial, if not too much time consuming, all the time. If there is some serious CPU overhead in Mesa (even those 8%) it should (in theory) make more massive difference in minimum FPS, or even maximum (similar case scenario - different reasons).

        Aerage framerate is as far away from indicative of performance, for example, what is better 200FPS AVG, 30 FPS 1% low, 500 FPS max, or 160FPS AVG, 50FPS 1% low, 370FPS MAX? Clearly 2nd option is better, that AVG FPS would never show, now for sure, that's (maybe?) not something that will happen so often, or maybe ever, and I don't think that Mesa have better lows etc., but still for the sake of accuracy, it would be nice to have those tests in all software (non-selective, for the sake of avoiding cherry picking) covered from time to time.

        Comment


        • #14
          This ends up being the same as GLXGears is not a benchmark... if you want to say the driver is CPU bound... you also need to include a CPU utilization graph, and preferably of the game vs driver utilization also.... it could be other factors causing the slightly slower numbers on AMD cards than just the driver being CPU bound... for instance it could be that AMD just isn't page flipping as fast or something like that... which is pretty irrelevant as long as they can do it at 144+fps since very very few people every run monitors faster than that hz.

          Its like sure GLXgears runs at 1000+fps.... this doesn't acutally mean *anything*

          Comment


          • #15
            Originally posted by cb88 View Post
            Its like sure GLXgears runs at 1000+fps.... this doesn't acutally mean *anything*
            If high numbers in a window does not satisfy you then run glxgears with fullscreen switch at 8K resolution, it might start to mean something
            Last edited by dungeon; 22 May 2017, 12:38 PM.

            Comment


            • #16
              Originally posted by dungeon View Post

              If high numbers in a window does not satisfy you then run glxgears with fullscreen switch at 8K resolution, it might start to mean something
              Yeah it tells you if your card can even paint the screen fast enough... but you probably already know that so yeah not useful.

              Comment


              • #17
                Originally posted by cb88 View Post
                Yeah it tells you if your card can even paint the screen fast enough... but you probably already know that so yeah not useful.
                For developers everything should be useful, even glxears results. For users likely no, since it is not real world app of particular interests

                But could be useful even for users as coil whine test, sometimes code can introduce coil whine so you can even bisect code by listening to that
                Last edited by dungeon; 22 May 2017, 12:57 PM.

                Comment


                • #18
                  Originally posted by marek View Post
                  OMG it's so close!
                  When you would beat Nvidia in this case you might even beat Windows at the same time.

                  Comment


                  • #19
                    Originally posted by oooverclocker View Post
                    When you would beat Nvidia in this case you might even beat Windows at the same time.
                    Not that long time ago in a galaxy far, far away... look at a TF 2 results it is nVidia diff

                    In whole, if you compare AMD Linux vs Windows on TF2, you would probably realise it runs at 66.6% of Windows performance, so to not beat but just to equal to Windows we need another 33.33% more perf from somewhere
                    Last edited by dungeon; 22 May 2017, 01:39 PM.

                    Comment


                    • #20
                      Well, there's always Tim's KHR_NO_ERROR work that's in progress. I wonder if enabling that shows any performance boost or not.

                      Comment

                      Working...
                      X