Announcement

Collapse
No announcement yet.

Open-Source RadeonSI Gallium3D vs. AMD Catalyst On Linux

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

  • Open-Source RadeonSI Gallium3D vs. AMD Catalyst On Linux

    Phoronix: Open-Source RadeonSI Gallium3D vs. AMD Catalyst On Linux

    Towards the end of June I published AMD RadeonSI Gallium3D benchmarks, the open-source Linux graphics driver supporting the Radeon HD 7000/8000 series hardware on Linux. While the alternative to the Catalyst driver can accelerate OpenGL, it's very slow. Open-source driver benchmarks were shown in that article compared to older generations of AMD Radeon hardware backed by the mature R600 Gallium3D driver. In this article are benchmarks comparing the open-source "RadeonSI" driver to the proprietary AMD Catalyst GPU driver on the Radeon HD 7850/7950 graphics cards. As an additional driver reference point were also Radeon HD 7950 Cayman results; all testing happened from Fedora 19 Linux.

    http://www.phoronix.com/vr.php?view=18870

  • #2
    I'm not too surprised with these results. My 7850 defaults to 180Mhz on the 3.8-3.10 kernels, but runs at 860Mhz on 3.11 with the new dynamic power management code. There is a similar gap in memory speeds as well.

    Once there are tests running on that newer kernel, the results should improve somewhat. There's definitely still work to do, but the gap should shrink a bit.

    Comment


    • #3
      What's the sense in benchmarking 3.9 while 3.10 is already stable? Nonsense.
      ## VGA ##
      AMD: X1950XTX, HD3870, HD5870
      Intel: GMA45, HD3000 (Core i5 2500K)

      Comment


      • #4
        Yes, this is stupid. We know performance still sucks with older kernels. 3.10 finally introduced tiling which likely improves performance a lot. 3.11 has even more improvements, like DPM (which finally allows proper reclocking) and PCIe gen2/gen3 support for radeonsi. Please make sensible benchmarks.

        Comment


        • #5
          Originally posted by darkbasic View Post
          What's the sense in benchmarking 3.9 while 3.10 is already stable? Nonsense.
          It's just a guess, but I'm guessing that 3.9 is just what Fedora 19 shipped with.

          Comment


          • #6
            Originally posted by Veerappan View Post
            It's just a guess, but I'm guessing that 3.9 is just what Fedora 19 shipped with.
            Yes, Fedora doesn't have 3.10 yet, its still 3.9.8

            Michael some notes....

            You said default configurations, correct? That means medium power profile, most likely, so some of these results aren't fair given that Catalyst as full power management and can run at max clocks as needed, but the open source driver defaults to medium. Also RadeonSI is by default crippled on clockspeed, it defaults to vbios settings if I'm not mistaken (And given a comment about, about the clock difference, it appears Im right).

            I am glad to see that R600g is basically at or close to Catalyst, and I thank you for that. But personal opinion: wait until you are running the new DPM work to do more Radeon articles. THATS when things will be a lot more fair to the open source drivers because clock speeds will be similar.

            Comment


            • #7
              Originally posted by Ericg View Post
              Yes, Fedora doesn't have 3.10 yet, its still 3.9.8

              Michael some notes....

              You said default configurations, correct? That means medium power profile, most likely, so some of these results aren't fair given that Catalyst as full power management and can run at max clocks as needed, but the open source driver defaults to medium. Also RadeonSI is by default crippled on clockspeed, it defaults to vbios settings if I'm not mistaken (And given a comment about, about the clock difference, it appears Im right).

              I am glad to see that R600g is basically at or close to Catalyst, and I thank you for that. But personal opinion: wait until you are running the new DPM work to do more Radeon articles. THATS when things will be a lot more fair to the open source drivers because clock speeds will be similar.
              Its not about fairness.

              And Your notion of "fairness" is strange as You ignore that it was AMD team decision to make those defaults You find "unfair"...

              It was about what Linux users can expect. And they most likely run on 3.9 or older kernels... Since no mainstream distro (mobile or not) use 3.10 yet. Such test would describe future for us.

              Comment


              • #8
                Originally posted by przemoli View Post
                It was about what Linux users can expect.
                It's about what Linux users can expect if they don't install the latest driver.

                Compared to a recent version of the closed driver which must be downloaded from AMD's webpage and for which you have to compile a kernel module.

                I suggest that it would be more interesting to see what Linux drivers can do now. I mean, it is rather widespread in the benchmarking community that you install drivers before you benchmark new hardware.

                And Your notion of "fairness" is strange as You ignore that it was AMD team decision to make those defaults You find "unfair"...
                It was also AMD team's decision to release drivers which can reclock such GPUs dynamically.

                So it is indeed odd that this test does not use drivers which are supposed to be used with such hardware. I.e. those found in the current 3.11 Linux tree.
                Last edited by pingufunkybeat; 07-05-2013, 03:21 PM.

                Comment


                • #9
                  Originally posted by pingufunkybeat View Post
                  It's about what Linux users can expect if they don't install the latest driver.

                  Compared to a recent version of the closed driver which must be downloaded from AMD's webpage and for which you have to compile a kernel module.

                  I suggest that it would be more interesting to see what Linux drivers can do now. I mean, it is rather widespread in the benchmarking community that you install drivers before you benchmark new hardware.


                  It was also AMD team's decision to release drivers which can reclock such GPUs dynamically.

                  So it is indeed odd that this test does not use drivers which are supposed to be used with such hardware. I.e. those found in the current 3.11 Linux tree.
                  FLOSS drivers do not come in form of kernel module that can work with multiple kernel versions. Period. And benchmarking community can use OpenBenchmarking.com to share their tests.
                  (Also I forgot to meantion that we do not know when those test where performed. First impresion of "imediatly" may be wrong.)

                  Comment


                  • #10
                    Well, if anything, this will be a good reference point to see how the performance evolves in later kernels.

                    Comment


                    • #11
                      Originally posted by przemoli View Post
                      Its not about fairness.

                      And Your notion of "fairness" is strange as You ignore that it was AMD team decision to make those defaults You find "unfair"...

                      It was about what Linux users can expect. And they most likely run on 3.9 or older kernels... Since no mainstream distro (mobile or not) use 3.10 yet. Such test would describe future for us.
                      Well, i do find it strange that a user concerned about performance wouldn't even bother setting the performance profile setting to high.

                      I know there are some issues with that on APUs, at least, (not sure about SI) - but it would likely help at least SOME... And it takes about 5 seconds to do, even on the 3.9 kernel. No recompiling or installing necessary.

                      But i agree this is a nice data point to have to compare against in future tests. It's just that those future tests are what will really be interesting, instead of this one.

                      Comment


                      • #12
                        Just made a quick test on Fedora 19 with my 7750, using Nexuiz at 1920x1200. (nexuiz-linux-x86_64-glx -benchmark demos/demo1 -nosound)

                        After bootup
                        Code:
                        default engine clock: 800000 kHz
                        current engine clock: 299990 kHz
                        default memory clock: 1125000 kHz
                        current memory clock: 148990 kHz
                        voltage: 1100 mV
                        
                        1910 frames 95.9403298 seconds 19.9082076 fps, one-second fps min/avg/max: 13 21 33 (90 seconds)
                        1910 frames 96.4965858 seconds 19.7934464 fps, one-second fps min/avg/max: 13 20 31 (90 seconds)
                        After setting /sys/class/drm/card0/device/power_profile to high
                        Code:
                        default engine clock: 800000 kHz
                        current engine clock: 799990 kHz
                        default memory clock: 1125000 kHz
                        current memory clock: 1124990 kHz
                        voltage: 1100 mV
                        
                        1910 frames 41.9272680 seconds 45.5550788 fps, one-second fps min/avg/max: 30 48 61 (90 seconds)
                        1910 frames 42.0657239 seconds 45.4051380 fps, one-second fps min/avg/max: 30 47 61 (90 seconds)

                        Comment


                        • #13
                          As others have said, RadeonSI performance can be greatly improved either by selecting profile based power management and then setting /sys/class/drm/card0/device/power_profile to high, or by using a 3.11 kernel. Interestingly, when using a 3.11 kernel it seems like it will not be necessary to enable the new "DPM" power mangement (by adding radeon.dpm=1 to the GRUB kernel boot options) to get good performance (although of course enabling DPM will reduce power consumption and lower heat output) - I have found the following patch in the 3.11 kernel:

                          author Alex Deucher 2013-07-05 17:14:30 (GMT)
                          committer Alex Deucher 2013-07-05 22:08:54 (GMT)
                          commit c6cf7777a32da874fabec4fd1c2a579f0ba4e4dd
                          tree 22a8b1f3b98714760a24b69f7d45d56c716dcfe0
                          parent 338a95a95508537e23c82d59a2d87be6fde4b6ff

                          drm/radeon: set default clocks for SI when DPM is disabled

                          Fix patching of vddc values for SI and enable manually forcing clocks to default levels as per NI.

                          This improves the out of the box performance with SI asics.

                          Signed-off-by: Alex Deucher

                          Comment

                          Working...
                          X