Announcement

Collapse
No announcement yet.

10-Way NVIDIA GeForce GTX OpenCL & CUDA Performance Benchmarks

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

  • 10-Way NVIDIA GeForce GTX OpenCL & CUDA Performance Benchmarks

    Phoronix: 10-Way NVIDIA GeForce GTX OpenCL & CUDA Performance Benchmarks

    With having just added some new OpenCL/CUDA benchmarks to the Phoronix Test Suite and OpenBenchmarking.org, I took this opportunity to run a variety of OpenCL/CUDA GPGPU tests on a wide-range of NVIDIA GeForce graphics cards.

    http://www.phoronix.com/vr.php?view=22388

  • #2
    Originally posted by phoronix View Post
    Phoronix: 10-Way NVIDIA GeForce GTX OpenCL & CUDA Performance Benchmarks

    With having just added some new OpenCL/CUDA benchmarks to the Phoronix Test Suite and OpenBenchmarking.org, I took this opportunity to run a variety of OpenCL/CUDA GPGPU tests on a wide-range of NVIDIA GeForce graphics cards.
    http://www.phoronix.com/vr.php?view=22388
    The table in the article is showing invalid frequencies for GTX 950: 135/405MHz

    Comment


    • #3
      Originally posted by << ⚛ >> View Post

      The table in the article is showing invalid frequencies for GTX 950: 135/405MHz
      It's likely just some fallback being used by the NVCTL extension since I don't recall if the GTX 950 was "officially" supported then or not, etc.
      Michael Larabel
      http://www.michaellarabel.com/

      Comment


      • #4
        Originally posted by phoronix View Post
        Phoronix: 10-Way NVIDIA GeForce GTX OpenCL & CUDA Performance Benchmarks

        With having just added some new OpenCL/CUDA benchmarks to the Phoronix Test Suite and OpenBenchmarking.org, I took this opportunity to run a variety of OpenCL/CUDA GPGPU tests on a wide-range of NVIDIA GeForce graphics cards.

        http://www.phoronix.com/vr.php?view=22388
        "SHOC: CUDA - Texture Read Bandwidth" is flawed.

        Comment


        • #5
          Originally posted by << ⚛ >> View Post

          "SHOC: CUDA - Texture Read Bandwidth" is flawed.
          The upstream code doesn't look like it scales too well for that sub-test, is that what you mean?
          Michael Larabel
          http://www.michaellarabel.com/

          Comment


          • #6
            Is nVIDIA GeForce GTX 750 Ti a good card for using Blender when using Cycles renderer? I have an AMD A8-7600 APU which I cannot use OpenCL for basic 3D rendering.

            Consider that my video (between 0:00 to 0:37) would take 2 days to render:
            https://www.youtube.com/watch?v=B0C43BOxMpU

            But note, however, that I used Blender's internal renderer and not Cycles renderer. Plus, this is for my video project that I did two years ago.

            I want a video card with a maximum of 75W TDP.

            Comment


            • #7
              Originally posted by Michael View Post
              The upstream code doesn't look like it scales too well for that sub-test, is that what you mean?
              I looked into the code of SHOC and cannot explain why "CUDA - Texture Read Bandwidth" would generate such results for all GTX 9xx.

              Maybe those results are correct and represent an architectural limit of all GTX 9xx GPUs when reading textures from GPU cache - though I find this hard to believe.

              Another option would be that CUDA 7.5 compiler applied some strange optimization to the code.

              Comment


              • #8
                Originally posted by GraysonPeddie View Post
                Is nVIDIA GeForce GTX 750 Ti a good card for using Blender when using Cycles renderer? I have an AMD A8-7600 APU which I cannot use OpenCL for basic 3D rendering.
                Frame 0 of ring_27.blend (http://download.blender.org/demo/test/Demo_274.zip):

                [email protected]: 2m05s
                GTX 750 cuda: 1m47s

                Comment


                • #9
                  Originally posted by siavashserver
                  Also your APU weirdly runs fast
                  Not really My i7-4750HQ 2Ghz Notebook CPU renders this scene in 1m11s.

                  But it is pretty annoying that Blender does not support OpenCL that well. I mean, yeah AMDs OpenCL Compiler is shit but LuxRender seems to handle that pretty well.

                  At least under Windows AMD Cards are beating the shit out of NVidia Cards in LuxMark.

                  Comment


                  • #10
                    Originally posted by siavashserver
                    Would you mind trying your benchmark with bigger tile sizes? As far as I know GPUs don't like tiny (32px) render tile sizes. Also your APU weirdly runs fast Here are my results when running with different tile sizes:

                    (Running on Windows 10 x64, Blender 2.76b x64, Core2 Quad Q9550 2.8GHz, 4GB DDR2 memory, GTX960 4GB)

                    https://www.meta-chart.com/share/rin...9550-vs-gtx960
                    [email protected] cpu: 2m05s (tile=32x32), 2m03s (16x16)
                    GTX 750 cuda: 1m47s (tile=32x32), 54s (64x64), 44s (128x128), 43.89s (256x256)

                    Comment

                    Working...
                    X