Announcement

Collapse
No announcement yet.

NVIDIA Performance: Windows vs. Linux vs. Solaris

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

  • NVIDIA Performance: Windows vs. Linux vs. Solaris

    Phoronix: NVIDIA Performance: Windows vs. Linux vs. Solaris

    Earlier this week we previewed the Quadro FX1700, which is one of NVIDIA's mid-range workstation graphics cards that is based upon the G84GL core that in turn is derived from the consumer-class GeForce 8600 series. This PCI Express graphics card offers 512MB of video memory with two dual-link DVI connections and support for OpenGL 2.1 while maintaining a maximum power consumption of just 42 Watts. As we mentioned in the preview article, we would be looking at this graphics card's performance not only under Linux but also testing this workstation solution in both Microsoft Windows and Sun's Solaris. In this article today, we are doing just that as we test the NVIDIA Quadro FX1700 512MB with each of these operating systems and their respective binary display drivers.

    http://www.phoronix.com/vr.php?view=11968

  • #2
    Image quality in Linux might not be up to par

    NVIDIA might have cut some corners when it comes to image quality on Linux which could explain why it seems faster than in Vista.

    Comment


    • #3
      Originally posted by stan View Post
      NVIDIA might have cut some corners when it comes to image quality on Linux which could explain why it seems faster than in Vista.
      Nope. Vista's as much as 20-40% slower than XP. That's just a fact on either NVidia or AMD. Vista's design doesn't allow for as good a performance as previous versions of Windows, just so you know.

      There's a whole host of things that're going on and unless you've got proof of degraded quality, don't suppose this out of the gate.

      Comment


      • #4
        Excellent review guys. SPECViewPerf is a standard benchmark for me and the industry I work in (I am the sysadmin at a small studio called Kanuka).

        Currently Kubuntu Linux on NVidia hardware is the standard workstation rollout for Kanuka, with a few MacOSX machines and one WindowsXP Pro box for 3DSMax work. Vista has been completely shunned from the studio, and for good reason as these benchmarks show.

        This review mirrors a lot of our own in-house testing, and reaffirms for us that we made the right decision not only for flexibility reasons, but also for performance reasons.

        Comment


        • #5
          Sweet news.

          I'm going to try that benchmark myself too, it looks quite interesting.

          Comment


          • #6
            Originally posted by Vadi View Post
            Sweet news.

            I'm going to try that benchmark myself too, it looks quite interesting.
            It's something that's pretty aggressive and will really beat on your GPU hard in ways that games won't do.

            Comment


            • #7
              You were right about the aggressive part - it was raping my laptop for half hour, and it was making quite a bit of weird squeaky noises.

              I did notice that it was using both of my cpu's though, each at half-power - is my setup messed up, or was it really supposed to use the CPU?

              Comment


              • #8
                Originally posted by Vadi View Post
                You were right about the aggressive part - it was raping my laptop for half hour, and it was making quite a bit of weird squeaky noises.

                I did notice that it was using both of my cpu's though, each at half-power - is my setup messed up, or was it really supposed to use the CPU?
                OpenGL is single threaded (indeed, multi-threading any realtime API is damn near impossible).

                The benchmark is designed to stress your GPU (video card), but with that said it will need the CPU to feed it the data it needs. Modern video cards are still fairly primitive in what they can process. While they have been offloading "Transformation and Lighting" effects, and can now do pixel shaders and whatnot, most geometry calculation and texture preparation is still done by the host system in software, and then fed to the GPU at a later time to do the final rasterization to the output framebuffer, and then finally your screen.

                What you saw was the single thread jumping back and forth between your two processors, which is not all that uncommon on any multiprocessor system. And yes, it's perfectly normal behaviour for SPECViewPerf. Most of the models in SPECViewPerf have much higher poly counts (some by several orders of magnitude) than anything you'll find in a modern video game. As such, the CPU must work a lot harder to feed the GPU the information it needs to draw the scene.
                Last edited by elvis; 03-08-2008, 05:14 PM.

                Comment


                • #9
                  Oh, okey.

                  I still have yet to graph this data, but I did get two things - a) starting games on a separate X server (log out, switch session to failsafe terminal, start game) can be put into the "compiling your own kernel" category - useless on newer hardware, since the difference is minimal and not even clear, and b) same thing for compiz - it looks like there's no opengl & compiz conflict on the nvidia card, games run just fine, and there was not much diff with compiz on/off.

                  Here's my raw data if anyone wants. I did it with two drivers, 169.07 and 169.12. Graphics card is "GeForce 8600M GT", CPU "Intel(R) Core(TM)2 Duo CPU T5250 @ 1.50GHz":

                  169.07, with metacity, try 1:

                  Run All Summary
                  3dsmax-04 Weighted Geometric Mean = 10.20
                  catia-02 Weighted Geometric Mean = 10.55
                  ensight-03 Weighted Geometric Mean = 11.43
                  light-08 Weighted Geometric Mean = 9.204
                  maya-02 Weighted Geometric Mean = 25.62
                  proe-04 Weighted Geometric Mean = 8.459
                  sw-01 Weighted Geometric Mean = 16.50
                  ugnx-01 Weighted Geometric Mean = 3.879
                  tcvis-01 Weighted Geometric Mean = 5.162


                  169.07, with metacity, try 2 (completely left the desktop alone):

                  Run All Summary
                  3dsmax-04 Weighted Geometric Mean = 10.01
                  catia-02 Weighted Geometric Mean = 10.48
                  ensight-03 Weighted Geometric Mean = 11.40
                  light-08 Weighted Geometric Mean = 9.073
                  maya-02 Weighted Geometric Mean = 25.83
                  proe-04 Weighted Geometric Mean = 8.569
                  sw-01 Weighted Geometric Mean = 16.54
                  ugnx-01 Weighted Geometric Mean = 3.898
                  tcvis-01 Weighted Geometric Mean = 5.210


                  169.07, failsafe terminal, black screen kicked in twice:

                  Run All Summary
                  3dsmax-04 Weighted Geometric Mean = 10.40
                  catia-02 Weighted Geometric Mean = 10.55
                  ensight-03 Weighted Geometric Mean = 11.51
                  light-08 Weighted Geometric Mean = 8.417
                  maya-02 Weighted Geometric Mean = 26.01
                  proe-04 Weighted Geometric Mean = 8.533
                  sw-01 Weighted Geometric Mean = 16.50
                  ugnx-01 Weighted Geometric Mean = 3.869
                  tcvis-01 Weighted Geometric Mean = 5.126


                  169.12, with metacity, try 1 (completely left the desktop alone):

                  Run All Summary
                  3dsmax-04 Weighted Geometric Mean = 10.11
                  catia-02 Weighted Geometric Mean = 10.72
                  ensight-03 Weighted Geometric Mean = 11.42
                  light-08 Weighted Geometric Mean = 8.963
                  maya-02 Weighted Geometric Mean = 25.72
                  proe-04 Weighted Geometric Mean = 8.423
                  sw-01 Weighted Geometric Mean = 16.59
                  ugnx-01 Weighted Geometric Mean = 3.891
                  tcvis-01 Weighted Geometric Mean = 5.211

                  169.12, with metacity, try 2 (completely left the desktop alone):

                  Run All Summary
                  3dsmax-04 Weighted Geometric Mean = 9.892
                  catia-02 Weighted Geometric Mean = 10.54
                  ensight-03 Weighted Geometric Mean = 11.54
                  light-08 Weighted Geometric Mean = 8.991
                  maya-02 Weighted Geometric Mean = 25.59
                  proe-04 Weighted Geometric Mean = 8.455
                  sw-01 Weighted Geometric Mean = 16.53
                  ugnx-01 Weighted Geometric Mean = 3.889
                  tcvis-01 Weighted Geometric Mean = 5.157


                  169.12, with compiz, try 1 (completely left the desktop alone):

                  Run All Summary
                  3dsmax-04 Weighted Geometric Mean = 10.38
                  catia-02 Weighted Geometric Mean = 9.773
                  ensight-03 Weighted Geometric Mean = 10.73
                  light-08 Weighted Geometric Mean = 9.119
                  maya-02 Weighted Geometric Mean = 25.37
                  proe-04 Weighted Geometric Mean = 8.387
                  sw-01 Weighted Geometric Mean = 16.13
                  ugnx-01 Weighted Geometric Mean = 3.714
                  tcvis-01 Weighted Geometric Mean = 5.087

                  169.12, with compiz, try 2 (completely left the desktop alone):

                  Run All Summary
                  3dsmax-04 Weighted Geometric Mean = 9.997
                  catia-02 Weighted Geometric Mean = 10.49
                  ensight-03 Weighted Geometric Mean = 10.72
                  light-08 Weighted Geometric Mean = 9.179
                  maya-02 Weighted Geometric Mean = 25.26
                  proe-04 Weighted Geometric Mean = 8.328
                  sw-01 Weighted Geometric Mean = 16.10
                  ugnx-01 Weighted Geometric Mean = 3.716
                  tcvis-01 Weighted Geometric Mean = 5.112

                  Comment


                  • #10
                    Slashdotted!

                    http://hardware.slashdot.org/article.../03/09/1958223

                    Comment


                    • #11
                      Originally posted by Rhettigan View Post
                      Oh, my... FRESHLY Slashdotted...

                      Best of luck, Michael- just be glad it's Sunday and not Monday that they did this one on... >:-)

                      Comment


                      • #12
                        Originally posted by Svartalf View Post
                        Oh, my... FRESHLY Slashdotted...

                        Best of luck, Michael- just be glad it's Sunday and not Monday that they did this one on... >:-)
                        Haven't got taken down by a SlashDot in about a year... I am not worried at all. In recent months the servers (and the Phoronix Content Management System) have handled Digg, SlashDot, Engadget, etc fine Heck Phoronix has had 60+ Digg front-pages, and I think only on three or four cases did it bring the server down.
                        Michael Larabel
                        http://www.michaellarabel.com/

                        Comment


                        • #13
                          The SPECViewperf 9.0.3 results for Solaris Express Developer
                          Edition 1/08 are low for two reasons.

                          First, when adding xVM dom0 support to the NVIDIA graphics
                          driver I broke a path in the kernel driver which optimizes
                          OpenGL performance. This is being tracked under Sun bug:

                          6665672: fix for 6593812 causes OpenGL performance regression

                          This will be fixed in Solaris Express Community Edition
                          build 85.


                          Second, to clarify an earlier comment by elvis, while the
                          SPECViewperf 9.0.3 benchmark is single threaded, OpenGL itself
                          has MT support. (See section 3.7 of the OpenGL ABI for Linux).
                          The Linux driver runs in single thread mode if it does not
                          detect the application was built with -lpthread. Solaris
                          has a unified thread model which forces the driver into
                          thread safe mode by default. Since Viewperf 9.0.3 is compiled
                          without -lpthread, by default it runs in faster single thread
                          mode on Linux and thread safe mode on Solaris. To run
                          Viewperf 9.0.3 equivalent to Linux single threaded mode,
                          set the environment variable:

                          export __GL_SINGLE_THREADED=1

                          I don't expect this issue will ever be "fixed".

                          Comment


                          • #14
                            Originally posted by Michael View Post
                            Haven't got taken down by a SlashDot in about a year... I am not worried at all. In recent months the servers (and the Phoronix Content Management System) have handled Digg, SlashDot, Engadget, etc fine Heck Phoronix has had 60+ Digg front-pages, and I think only on three or four cases did it bring the server down.
                            Heh... I'd not get my server plant taken down, but I'd sure as heck have the pipe choked off if I got a good old-fashioned Slashdotting.

                            Comment

                            Working...
                            X