Announcement

Collapse
No announcement yet.

What Linux Benchmarks Would You Like To See Next?

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

  • #51
    More OpenCL/GPGPU software benchmarks

    I would like to see more benchmarks for OpenCL/GPGPU software (especially scientific).
    Comparison between latest GeForce 980 and rest of the graphics card will be good idea.
    Another thing that would be in benchmarks is comparison between single precision and double precision performance.
    Comparison between professional graphics cards (Quadro and FirePro) and gaming hardware also will be appriecated.

    Comment


    • #52
      Wine benchmarks with Gallium-Nine vs CSMT vs vanilla.

      Comment


      • #53
        Originally posted by dungeon View Post
        Why is that interesting? All drivers have different memory managment, nothing to conclude there about blobs . Only opensource ones have default, but you can change it
        It's extremely interesting if Michael's data can point out app Z is bottlenecked by YPU on card X, but not on the blob. That's the kind of data that only multi-card benchmarking can provide.

        Also the reverse, if the blob is bottlenecked but open driver not, that would be a great result.

        Comment


        • #54
          Originally posted by linuxgeex View Post
          @Phoronixria:

          Yes, benching Gimp is actually easy since it's scriptable. So is Firefox for that matter. Both perform much better on 64-bit distros, and in some cases with hardware acceleration enabled. These numbers are meaningful to a lot of users!
          Any example Gimp scripts then for benchmarking?
          Michael Larabel
          https://www.michaellarabel.com/

          Comment


          • #55
            Blender/Cycles as a benchmark

            Hello Michael, I had some conversations with Blender developers about using Cycles as a OpenCL/CUDA benchmark.

            Lukas T?nne
            According to Lukas using Cycles standalone is possible but not practical, it needs a special xml file with all (binary) data encoded inside. Using Cycles linked with Blender allows rendering blend files directly.

            Cycles should be deterministic, rendering the same scene with the same version and configuration should give pixel for pixel the same result. (I (Tim) don't know if this is also true for metadata in output like maybe timestamps?)

            Interesting tests for Blender would be stress testing drivers. AMD binary drivers are notoriously bad as it has no real support for functions, it essentially duplicates all code in memory, the Cycles GPU kernel grows exponentially with these drivers. See http://wiki.blender.org/index.php/De.../Cycles/OpenCL
            Nvidia's binary divers are mature and Cycles works best with these (CUDA) drivers. Still Blender developers want Cycles to work with any configuration. The problem is that they are very dependent on driver development for this, so anything that pushes hardware manufacturers to improve drivers is welcome.
            Next to that, testing for improvements in render speed and more specific, improvements in gpu memory access would be interesting for Blender.

            Best would be to use several different scenes that exercise the different functions of the Cycles renderer, specifically hair rendering.

            If anyone (Michael) has any questions they can drop by at the following mailing-list and Lukas or other developers would be more then willing to answer them: http://lists.blender.org/mailman/listinfo/bf-committers

            Sergey Sharybin
            Sergey also thinks it's easier to use Cycles with Blender.

            The Cycles API gives information about the time it takes to render a frame, this data can be accessed within Blender with Python scripting. It would be possible to extend the API to provide other metrics.

            According to Sergey Cycles has no OpenMP support (I (Tim) do recall a OpenMP option somewhere..)

            Sergey also points to the bf-committers mailing-list for any questions.

            Campbell Barton
            According to Cambell it's perfectly possible to use Cycles as a benchmark. You can "just do it".

            Comment


            • #56
              APU next to gpu

              I would like that APU were benchmarked next with gpu, so one can compare relative performance. Even if cpu is different, make one test with both the cpu and the apu with the same card (one more or less stable) and then with the apu "gpu". Example: test the radeon cards and nvidia cards with a intel i7, then pick the R9 270x and benchmark it with the A10 (so we can compare the change due the cpu) and then benchmark the A10 APU directly. If you add a intel HD, do the same thing, one test with a normal gpu card, other with the integrated gpu

              Comment


              • #57
                When Broadwell comes out, I'd like to see graphics benchmarks for single channel vs. dual channel vs. "flex mode". It seems there's quite a few laptops (Asus, Lenovo) these days that come soldered with 4GB and have 1 free DIMM slot. So upgrading to 12GB ends up in an asymmetric configuration. Intel's documentation says that in flex mode, the lower part of memory (e.g. 4 GB from each channel) are accessed in dual-channel mode, and the remainder is accessed in single-channel. Given that onboard graphics are pretty constrained by memory bandwidth, I'd like to see if there is a negative impact of running in flex mode. If the CPU is smart enough to keep the video memory in the dual-channel area, then I suppose it shouldn't be an issue, but I haven't seen any tests done to show this. So I'd be interested in knowing this before my next laptop purchase when Broadwell comes along...

                Comment

                Working...
                X