Announcement

Collapse
No announcement yet.

Nouveau Drags Behind Intel & Radeon For Linux 2D Performance

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

  • Nouveau Drags Behind Intel & Radeon For Linux 2D Performance

    Phoronix: Nouveau Drags Behind Intel & Radeon For Linux 2D Performance

    While 3D/OpenGL is our primary focus of performance tests when it comes to graphics cards on Linux, it's always interesting to go back and check on the 2D performance as it's still important for the Linux desktop experience. The 2D performance is becoming interesting right now as well due to Intel's driver defaulting to SNA and GLAMOR acceleration being tried by some drivers for faster 2D over OpenGL. In this article we have some fresh 2D benchmarks of Intel, NVIDIA, and AMD graphics hardware running an updated open-source GPU driver stack on Ubuntu 14.04 LTS.

    http://www.phoronix.com/vr.php?view=20034

  • #2
    Wrong benchmarks for Intel?

    In the article you say that the backend for the intel driver is SNA, but in the summary it says GLAMOR, which seems to correlate better with the numbers (in comparison with this article)

    Comment


    • #3
      Quick question... the article text suggests that the Intel driver is using SNA...

      The default acceleration methods for each of the drivers was used: Intel's 3.0 pre-release drivers default to SNA (rather than UXA on older series), Nouveau uses EXA, and the Radeon driver uses EXA by default for the Radeon HD 6000 series and older while the Radeon HD 7000 series and newer is limited to only supporting GLAMOR.
      ... but the screen grab above it suggests that the Intel driver is using GLAMOR. Am I misreading something ?

      Comment


      • #4
        I also runned this test on my system, and i found some strange results. A couple of times i have much better performance despite that i have:
        - 3.13 kernel
        - 10.0.2 mesa
        - A much older cpu (2500k)

        I think the 2D performance of glamor on intel is still not equivalent :=)

        http://openbenchmarking.org/result/1...SO-1403134PL86

        Comment


        • #5
          r600g results are suspiciously decent. I have a 6770 and it was way slower at running Aquaria (native build from source code) than Intel's ancient Q33 was on the same machine. I can't believe r600g beats Intel's SNA now!

          Comment


          • #6
            Originally posted by Shnatsel View Post
            r600g results are suspiciously decent. I have a 6770 and it was way slower at running Aquaria (native build from source code) than Intel's ancient Q33 was on the same machine. I can't believe r600g beats Intel's SNA now!
            Aquaria is an OpenGL game, and it has pretty much nothing to do with 2D acceleration tested here.

            Comment


            • #7
              Originally posted by Shnatsel View Post
              r600g results are suspiciously decent. I have a 6770 and it was way slower at running Aquaria (native build from source code) than Intel's ancient Q33 was on the same machine. I can't believe r600g beats Intel's SNA now!
              While beating SNA is quite a feat as of today, a 6770 is better than any of intel's GPUs so I think the sheer processing power of the GPU is why it happened to win. Keep in mind too that the radeon drivers have made tremendous progress. Ever since I ditched catalyst, I noticed I actually get as good or even better performance out of the open source drivers in most things (such as Portal or the Heaven benchmark). Some programs (such as Trine 2) perform a little worse, though not enough to complain. If I could go back in time 2-3 years ago and tell people that would happen, nobody would believe me.

              BTW, I own 2 HD5750s. I hope someone implements crossfire support in the near future; the extra performance I'd gain from that would "fix" the performance losses I get.


              Anyway, while nouveau obviously didn't do that great, I honestly expected worse. Considering the shortcomings and obstacles of the nouveau team, they've made better progress than what we give them credit for.

              Comment


              • #8
                Originally posted by Shnatsel View Post
                I have a 6770 and it was way slower at running Aquaria (native build from source code) than Intel's ancient Q33 was on the same machine.
                You can fill a bug at https://bugs.freedesktop.org Drivers/Gallium/r600 if that game is so slow... as i remember there is option in config file to turn off fbo maybe that can help with performance .

                Comment


                • #9
                  Originally posted by schmidtbag View Post
                  While beating SNA is quite a feat as of today, a 6770 is better than any of intel's GPUs so I think the sheer processing power of the GPU is why it happened to win. Keep in mind too that the radeon drivers have made tremendous progress. Ever since I ditched catalyst, I noticed I actually get as good or even better performance out of the open source drivers in most things (such as Portal or the Heaven benchmark). Some programs (such as Trine 2) perform a little worse, though not enough to complain. If I could go back in time 2-3 years ago and tell people that would happen, nobody would believe me.

                  BTW, I own 2 HD5750s. I hope someone implements crossfire support in the near future; the extra performance I'd gain from that would "fix" the performance losses I get.


                  Anyway, while nouveau obviously didn't do that great, I honestly expected worse. Considering the shortcomings and obstacles of the nouveau team, they've made better progress than what we give them credit for.
                  Assuming this earlier article which compares the different acceleration backends in the intel driver is correct, this one isn't - it's quite clearly the GLAMOR backend that has been benchmarked in this one. Case in point - look at the "GTK Widget: GtkDrawingArea - Circles" in both articles; with the same hardware SNA gets 10.13 in one article and 448.23 in the other...

                  Comment


                  • #10
                    Originally posted by bridgman View Post
                    Quick question... the article text suggests that the Intel driver is using SNA...



                    ... but the screen grab above it suggests that the Intel driver is using GLAMOR. Am I misreading something ?
                    From Xorg.0.log on the Intel machine (after digging through openbenchmarking.org):

                    (II) intel(0): Use GLAMOR acceleration.

                    So, no, you're not mistaken. The intel graphics option was also running glamor instead of SNA.

                    Comment


                    • #11
                      Originally posted by bridgman View Post
                      Quick question... the article text suggests that the Intel driver is using SNA...



                      ... but the screen grab above it suggests that the Intel driver is using GLAMOR. Am I misreading something ?
                      Whoops, might have merged the other Intel result with this multi-GPU 2D comparison... All the Intel SNA/UXA/GLAMOR results from the same system with testing the same day are in this result file http://www.phoronix.com/vr.php?view=19993 so will check into it then re-merge the results.
                      Michael Larabel
                      http://www.michaellarabel.com/

                      Comment


                      • #12
                        In 2012 Nouveau beat Radeon in Kdenlive performance

                        The Kdenlive video editor seems to involve one of the few tough XV video tests out there, as multiple video streams have to be resized for display in one "screen" portion of the Kdenlive GUI. It is common for playback to stutter during transitions when two files are played at once. Performance differs with different drivers: The proprietary drivers give the worst performance of all. Radeon gives good but not perfect performance-but the best, closest to stutter-free results I ever saw came from Nouveau running on an Nivdia GTS450 back in 2012. OpenGL performance was awful but nothing ever worked better in Kdenlive for some reason. I don't know if that's true today with all the general improvements I've seen in the R600 driver, however.

                        Comment


                        • #13
                          When it comes to testing nouveau please test the older, reclockable ones so we can get a 1-1.
                          Basically I'd like to know how well nouveau is actually written when it can reclock.

                          Comment


                          • #14
                            Originally posted by liam View Post
                            When it comes to testing nouveau please test the older, reclockable ones so we can get a 1-1.
                            Basically I'd like to know how well nouveau is actually written when it can reclock.
                            It's not really useful when the older drivers are on a different Gallium3D driver: NV50 vs. NVC0. And regardless of hardware generation the re-clocking on modern kernels requires a kernel rebuild at last check in order to bypass one of the checks... So for all intensive, real-world purposes for the vast majority of users, it's not easily re-clockable even for the older hardware that previously just required setting kernel parameters and writing some sysfs values.
                            Michael Larabel
                            http://www.michaellarabel.com/

                            Comment


                            • #15
                              Originally posted by Michael View Post
                              It's not really useful when the older drivers are on a different Gallium3D driver: NV50 vs. NVC0. And regardless of hardware generation the re-clocking on modern kernels requires a kernel rebuild at last check in order to bypass one of the checks... So for all intensive, real-world purposes for the vast majority of users, it's not easily re-clockable even for the older hardware that previously just required setting kernel parameters and writing some sysfs values.
                              Unless they changed something pretty recently you don't need to rebuild the kernel but you need to send the proper parameter to the kernel on boot (I don't recall it offhand since I made the change in etc/default many months ago).
                              The point about the driver split between generations is fair though.
                              From what I recall, when reclocking was available nouveau was surprisingly competitive. That's why I'm always interested in nv50.

                              Comment

                              Working...
                              X