Announcement

Collapse
No announcement yet.

RADV Vulkan Performance Appears To Improve With Linux 4.11

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

  • #11
    It might be fun to bisect and see what exactly caused this performance improvement.

    Comment


    • #12
      Originally posted by nanonyme View Post

      Are you asking about Vulkan or in general?
      Vulkan for starters and maybe DX9 gallium-nine vs. Nvidia (cause old games).

      Comment


      • #13
        Originally posted by Shevchen View Post

        Vulkan for starters and maybe DX9 gallium-nine vs. Nvidia (cause old games).
        It's not at all given thus far lower overhead in drivers leads to better overall performance. Shaders often aren't (and maybe shouldn't be?) optimized enough in games so drivers end up doing a degree of preprocessing removing redundant code. Driver overhead is primarily interesting when your CPU can't keep up with your GPU (so depends on individual machine build; granted, normally GPU isn't these days the bottleneck)

        Comment


        • #14
          Originally posted by davidbepo View Post

          then the r9 fury has a clear performance bug

          marek, any thoughts?
          It is not a bug. Let's look at 800x600:
          152.2 FPS / 127.3 FPS = 1.196
          1266 MHz / 1.000 MHz = 1.266

          So this tells us that one RX 480- shader is simply 26.6% faster and as a result the 480 as a whole is about 20% faster. This also tells us that the Fury can't benefit from more shaders in 800x600 in Dota 2 because you just can't split computing to an infinite amount of shaders, even with Vulkan.

          Let's have a look at 4K:
          53.97 FPS / 43.3 FPS = 1.246
          7.2 GFLOPS / 5.8 GFLOPS = 1.241

          So this just tells us that the Fury can utilize it's 3584 Shaders as well as the RX 480 can utilize it's 2304. -> Everything is alright.

          That's the reason why the Fiji chip is one of the worst balanced chips in my opinion. It's computing performance needs high details and big resolutions but it has only 4GiB VRAM.

          Comment


          • #15
            Originally posted by oooverclocker View Post

            It is not a bug. Let's look at 800x600:
            152.2 FPS / 127.3 FPS = 1.196
            1266 MHz / 1.000 MHz = 1.266

            So this tells us that one RX 480- shader is simply 26.6% faster and as a result the 480 as a whole is about 20% faster. This also tells us that the Fury can't benefit from more shaders in 800x600 in Dota 2 because you just can't split computing to an infinite amount of shaders, even with Vulkan.

            Let's have a look at 4K:
            53.97 FPS / 43.3 FPS = 1.246
            7.2 GFLOPS / 5.8 GFLOPS = 1.241

            So this just tells us that the Fury can utilize it's 3584 Shaders as well as the RX 480 can utilize it's 2304. -> Everything is alright.

            That's the reason why the Fiji chip is one of the worst balanced chips in my opinion. It's computing performance needs high details and big resolutions but it has only 4GiB VRAM.
            I am a happy 4k fury X gamer on radeonSI. I did have no problems with the 4GB VRAM in witcher 2, victor vran or cities skylines. Even in Hitman the normal Fury bashes the 980Ti at 4k despite the 980ti having 6GB, see http://www.phoronix.com/scan.php?pag...a-hitman&num=1

            Maybe Deus Ex is more demanding but I cannot find deus ex 4k benchmarks. Did Michael do those recently?

            From what I have read the 4GB is only a limit if you combine 4k with more than 2xAntiAliasing for most games. The stupid thing is, that at 4k you usually do not need AA anymore, but many benchmarks enable AA nevertheless because it is default on their ultra-settings, be it 4k or HD resolution.

            Comment


            • #16
              Originally posted by oooverclocker View Post

              It is not a bug. Let's look at 800x600:
              152.2 FPS / 127.3 FPS = 1.196
              1266 MHz / 1.000 MHz = 1.266

              So this tells us that one RX 480- shader is simply 26.6% faster and as a result the 480 as a whole is about 20% faster. This also tells us that the Fury can't benefit from more shaders in 800x600 in Dota 2 because you just can't split computing to an infinite amount of shaders, even with Vulkan.

              Let's have a look at 4K:
              53.97 FPS / 43.3 FPS = 1.246
              7.2 GFLOPS / 5.8 GFLOPS = 1.241

              So this just tells us that the Fury can utilize it's 3584 Shaders as well as the RX 480 can utilize it's 2304. -> Everything is alright.

              That's the reason why the Fiji chip is one of the worst balanced chips in my opinion. It's computing performance needs high details and big resolutions but it has only 4GiB VRAM.
              at 800x600 dota 2 is extremely cpu bound so there is clearly a bug

              Comment


              • #17
                Originally posted by tomtomme View Post
                I am a happy 4k fury X gamer on radeonSI. I did have no problems with the 4GB VRAM in witcher 2, victor vran or cities skylines.
                Yeah, with the games currently working on Linux more than 4GiB of VRAM are pretty much overkill, even in 4K. But also 4k Shaders are . So even though the Fury X has about 50% more raw performance you can mostly expect not that many more FPS than an RX 480 can do, except with very high resolutions and settings. Of course it's still a powerful card and you can be quite happy with the performance on Linux in pretty much all titles. For Windows AAA- titles of 2016, there are plenty of them that exceed 4GiB at least in 4K and hopefully we will see more of them on Linux in the future.

                That's just the main reason why AMD put so much effort in optimizing the utilization of shaders in Vega, as the top card in their lineup you just get the same amount of shaders like the Fury X has, just with higher clocks, better internal management and more VRAM. They could also have just shrunken Fiji and scaled it to 7K Shaders on a 600mm² die, running with 1.25GHz, resulting in 17.5 TFLOPS. But it would have been simply stupid because when the specific situation just utilizes 2500 shaders in average a GTX 1080 would be about 70% faster although it had just one third of the shader count. So shader clocks are also very important to compete with Nvidia cards in badly optimized and older games.
                That's just what I see in my opinion, when I look at several benchmarks and GPU specifications of the last years.

                Originally posted by davidbepo View Post
                at 800x600 dota 2 is extremely cpu bound so there is clearly a bug
                It doesn't really look CPU bound in the charts. The resolutions scale pretty reasonably with the FPS.
                Last edited by oooverclocker; 25 February 2017, 08:33 AM.

                Comment


                • #18
                  Edit:
                  Originally posted by tomtomme View Post
                  Maybe Deus Ex is more demanding but I cannot find deus ex 4k benchmarks.
                  You won't see less FPS when the VRAM is exceeded in contemporary games. There will just be some minor performance variations when the VRAM utilization hits about 90%. Then some data gets reallocated. Modern engines just change the textures with lower resolution textures or with a very blurry placeholder standard texture to prevent FPS drops. Only the quality of what you see will be worse.

                  Only when the VRAM is exceeded excessively so the reallocation process is initiated several times you will notice serious stuttering, very ugly textures/plopping textures and many variations in the low FPS.


                  This video shows the behavior, except that the guy who noticed it seems to believe that the VRAM isn't exceeded until it's completely occupied. If you know about the GTX 970- 3.5GiB- Bug, you know that the amount that is occupied in this video is extremely close to the maximum, especially when you are aware of the only way of working around this bug, which means pretty much limiting the maximum occupied amount to 3.5 GiB. Looking at the latest Resident Evil benchmarks the GTX 970 is sometimes even slower than a 960, which might be closely related. In that case, the textures don't just disappear or look ugly, the VRAM exists but is simply very slow without driver fixes.
                  Last edited by oooverclocker; 25 February 2017, 12:09 PM.

                  Comment

                  Working...
                  X