Announcement

Collapse
No announcement yet.

Mesa's RADV Vulkan Driver Holds A Narrowing Lead Over AMDVLK With Ubuntu 21.10 On Wayland

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

  • #11
    Ive only really ever used radv, too much effort to switch

    Comment


    • #12
      I think the real litmus test is Shadow Of The Tomb Raider. Certain games that can be extremely frequency sensitive I find not surprising, while its more a software optimization test those games that are frequency sensitive I can see pulling more frames with better drivers.

      Even when it comes to frame times it looks like splitting hairs.

      I find driver performance comparisons infinitely entertaining, and yes from time to time newer drivers come along that make things that much better for a lot of people.
      Last edited by creative; 30 December 2021, 12:51 PM.

      Comment


      • #13
        Originally posted by Masush5 View Post

        I've unscientifically timed it with shader pre-warming screens of some vulkan/d3d12 games and fossilize archive replays and open AMDVLK is still on average 10 times slower than RADV or the proprietary driver.
        We need specific tests with figures where progress is visible. "10 times slower" does not mean anything to the eye.

        Comment


        • #14
          Originally posted by user1 View Post
          It would be interesting to revisit RADV vs AMDVLK performance on Polaris, because last time I did some tests myself (before AMDVLK broke all my games with its switchable layer), the performance of AMDVLK was abysmal.
          Speaking of Polaris, async compute is apparently broken on all PAL based drivers, along with some crucial Vulkan extensions broken as well (that was finally fixed after 3 and a half months).
          Seems like AMD doesn't do enough QA for its PAL drivers on Polaris, which is still the most popular AMD architecture, at least according to Steam.
          Because Polaris was the last AMD GPU we could buy before Crypto and NFT drove prices to ridiculous levels. If you're like me and $200-300 for an RX 580 was a big purchase, you ain't gonna spend $500-600 on a 6600XT...pretty much the one GPU they've released that can be considered to be the upgrade path for the cheap-asses who stick to mid-range or lower cards.

          Comment


          • #15
            Originally posted by skeevy420 View Post

            Because Polaris was the last AMD GPU we could buy before Crypto and NFT drove prices to ridiculous levels. If you're like me and $200-300 for an RX 580 was a big purchase, you ain't gonna spend $500-600 on a 6600XT...pretty much the one GPU they've released that can be considered to be the upgrade path for the cheap-asses who stick to mid-range or lower cards.
            Well yeah.. I understand the current GPU situation very well. My point is that the QA of PAL drivers (DX12 and AMDVLK) seems lacking on Polaris, even though it's the most popular AMD architecture, so AMD should put more effort into it. I mean I haven't even heard about cases when Vulkan extensions just break in any other Vulkan driver.
            Last edited by user1; 30 December 2021, 02:13 PM.

            Comment


            • #16
              Originally posted by mphuZ View Post

              We need specific tests with figures where progress is visible. "10 times slower" does not mean anything to the eye.
              One specific test you can also do yourself is timing the initial shader compilation screen of the native vulkan game Detroit: Become Human, it has a free demo on steam. Make sure to clear all driver shader caches and also the game's own 'ShaderCache' folder it creates in it's install directory. The time it takes on my machine with RADV/ACO is 1min3s and with AMDVLK v-2021.Q4.3 it takes 8min55s, not quite 10x but close enough.
              I've seen similar results with Death Stranding (d3d12 game running on vkd3d-proton) which compiles all it's shaders on the first loading screen. Takes about 2 minutes 20 seconds with RADV and almost 20 minutes with AMDVLK, i've tried to redo the test with AMDVLK v-2021.Q4.3 but it seems like it also leaks memory now with pretty much all vkd3d-proton titles so it didn't finish.

              Comment


              • #17
                Originally posted by Seebi View Post

                Wow, that’s a large difference. Do you remember the games you tested? I’d be interested in testing as well.
                Also, when did you run these tests? I remember some improvements during the last months. LLVM still doesn’t reach the speeds of ACO, but the gap is smaller than before.
                A difference if 10x is most likely not caused by actual compilation speed, but due to shader caching (https://github.com/GPUOpen-Drivers/llpc/pull/1283 fixed a bug there, it was probably part of a release in October).
                See my comment above, and i always make sure to do these test with a clean driver cache so it's actual compilation time, not cache hits.

                Comment


                • #18
                  Originally posted by Masush5 View Post

                  One specific test you can also do yourself is timing the initial shader compilation screen of the native vulkan game Detroit: Become Human, it has a free demo on steam. Make sure to clear all driver shader caches and also the game's own 'ShaderCache' folder it creates in it's install directory. The time it takes on my machine with RADV/ACO is 1min3s and with AMDVLK v-2021.Q4.3 it takes 8min55s, not quite 10x but close enough.
                  I've seen similar results with Death Stranding (d3d12 game running on vkd3d-proton) which compiles all it's shaders on the first loading screen. Takes about 2 minutes 20 seconds with RADV and almost 20 minutes with AMDVLK, i've tried to redo the test with AMDVLK v-2021.Q4.3 but it seems like it also leaks memory now with pretty much all vkd3d-proton titles so it didn't finish.
                  I would like to see a test like this: https://gist.github.com/pendingchaos...089f29a8c6aa63
                  As you can see, even the ACO was inferior to the proprietary compiler in this test. I want to see the actual figures.

                  Comment


                  • #19
                    Originally posted by mphuZ View Post

                    I would like to see a test like this: https://gist.github.com/pendingchaos...089f29a8c6aa63
                    As you can see, even the ACO was inferior to the proprietary compiler in this test. I want to see the actual figures.
                    Don't you see that this is from June 2019 BEFORE ACO was even publicly announced and back then it supported only 2 or 3 shader types and the rest were still handled by LLVM, hence why Pro compiler is faster in these tests?
                    Seriously, I see you everywhere fanboying AMDVLK. Why? is it because it's the "official" Vulkan driver?

                    Comment


                    • #20
                      Originally posted by user1 View Post

                      Don't you see that this is from June 2019 BEFORE ACO was even publicly announced and back then it supported only 2 or 3 shader types and the rest were still handled by LLVM, hence why Pro compiler is faster in these tests?
                      Seriously, I see you everywhere fanboying AMDVLK. Why? is it because it's the "official" Vulkan driver?
                      Aren't you interested in seeing such a test?

                      Comment

                      Working...
                      X