Announcement

Collapse
No announcement yet.

RADV/RadeonSI Benchmarks On Mesa 18.0

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

  • #11
    Good news everyone... No regressions.

    Comment


    • #12
      Can you add mpv to your benchmarks? Their vulkan implementation is quite cutting edge and it's trivial to benchmark: https://haasn.xyz/posts/2017-10-05-h...hroughput.html

      Comment


      • #13
        Originally posted by GruenSein View Post

        Agreed. For example, on Windows Wolfenstein 2 is blazingly fast on Vegas now that it got optimized for it. Therefore, it is possible to extract the performance one would expect from these devices. The question is how much effort it requires on the developers part and how specific this is to Vega (it could just be the advancing Vulkan backend of the engine as well). If it requires lots of Vega-specific work I doubt that we will see many game developers putting in the hours...
        The new Wolfenstein seems to run relatively well on rx 580 cards too, so much of it it might be Amd specific optimizations.

        Comment


        • #14
          Good article. It may not be totally thorough, but it certainly has enough data to prove no major regressions.

          (Also I just wanted to mention that AMD's windows drivers are not at all related to mesa. You can't expect that a games behavior on windows will be anything at all like it's behavior in wine on linux. And then there's the issue of ported games, which many of them use middleware layers. So you can't expect the linux binary to behave anything at all like the windows binary. A lot of those middleware layers don't even try at all to emulate the behavior of windows libraries. They try to be 1 to 1 translations.)

          Comment


          • #15
            Nice benchmarks, good to see steady improvement. Just needs a little more Vulkan and Vega boost, although its already impressive.
            BTW, checked Unigine Valley yesterday on SI (HD 7970):
            Win10, dx11: 70fps
            Win10, opengl: 60fps
            Linux, mesa git: 65fps

            Comment


            • #16
              Originally posted by GruenSein View Post
              I find it somewhat concerning that a 10.5TFlops part like Vega 56 only performs marginally better than a 7.2TFlops Fury (non X) in most benchmarks. I am wondering if this is due immature drivers or bottle-necks in the fixed function block since compute performance quite good.
              Exactly what I was thinking

              Comment


              • #17
                Originally posted by GruenSein View Post
                I find it somewhat concerning that a 10.5TFlops part like Vega 56 only performs marginally better than a 7.2TFlops Fury (non X) in most benchmarks. I am wondering if this is due immature drivers or bottle-necks in the fixed function block since compute performance quite good.
                Nothing to do with Cpu. All Amd GPUs over 6-teraflops have flop to bandwidth ratio problems and the new Tile Renderer didn't solve this. Since all those products use HBM we don't now for sure if this mix is the problem. I don't suggest buying anything more than 6-teraflops from Amd, except if the price is a lot lower than the named flops. That way you invest in the future for half the price and for games that actually need this performance. Regardless of that i suggest undervolting any GCN graphics chip to 1.05v and overclock it to the maximum potential. As for HBM i propose overvolting and overclocking to a 90%, if you do this right for Vega 56 you will beat anything except a 1080ti. In the future you may get more.

                Comment


                • #18
                  Originally posted by artivision View Post

                  Nothing to do with Cpu. All Amd GPUs over 6-teraflops have flop to bandwidth ratio problems and the new Tile Renderer didn't solve this. Since all those products use HBM we don't now for sure if this mix is the problem. I don't suggest buying anything more than 6-teraflops from Amd, except if the price is a lot lower than the named flops. That way you invest in the future for half the price and for games that actually need this performance. Regardless of that i suggest undervolting any GCN graphics chip to 1.05v and overclock it to the maximum potential. As for HBM i propose overvolting and overclocking to a 90%, if you do this right for Vega 56 you will beat anything except a 1080ti. In the future you may get more.
                  Sounds logical but if that was the issue wouldn't we have the same kind of framerate ceiling on Windows?

                  Comment


                  • #19
                    Originally posted by IreMinMon View Post

                    Sounds logical but if that was the issue wouldn't we have the same kind of framerate ceiling on Windows?
                    They do have the same problem on windowz, probably even worse.

                    Comment

                    Working...
                    X