Announcement

Collapse
No announcement yet.

NVIDIA 24-Way GPU Comparison With Many OpenCL, CUDA Workloads

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

  • NVIDIA 24-Way GPU Comparison With Many OpenCL, CUDA Workloads

    Phoronix: NVIDIA 24-Way GPU Comparison With Many OpenCL, CUDA Workloads

    As part of re-testing all hardware prior to major GPU/driver launches, here is a look at the latest NVIDIA OpenCL/CUDA performance on Linux -- complementing the recent Blender 2.90 benchmarks and the latest NVIDIA vs. AMD Linux gaming performance. In still waiting to find out when we will get any NVIDIA Ampere hardware for Linux testing, I have been having some benchmarking fun and extended this to a 24-way graphics card comparison back to Maxwell in looking at not only the raw GPU compute performance but also the performance-per-Watt / power consumption and GPU thermal values.

    http://www.phoronix.com/vr.php?view=29492

  • #2
    percent results would be better..
    (good job)

    Comment


    • #3
      Originally posted by Niejaki View Post
      percent results would be better..
      (good job)
      You can click on the OpenBenchmarking.org link, then hit the "normalize results" checkbox. You can also use 'highlight result' options if wanting to normalize against a particular GPU. And other options on OB to slice and dice the data however you want.
      Michael Larabel
      http://www.michaellarabel.com/

      Comment


      • #4
        Michael, can you include also radeon via ROCm? OpenCL started working for Blender for navi and polaris. Polaris also woks in tensorflow.

        Comment


        • #5
          I was wondering, since there is lots of CUDA code around, did anybody of you try to port that to ROCm via the HIP compiler? AFAIK it was used for tensorflow and the performance is not bad. Quite often, the OpenCL code paths are not really good. Porting CUDA code could help, if this was "easy".

          Comment


          • #6
            But, … nobody really into OpenSource uses Nvidia anyway, ... :-/

            Comment


            • #7
              I would say from pespective of such productivity results, 2xxx serie is not anyhow failure, RTX 2060 is faster then GTX 1080. If you add new encoder/decoder stuff, it was success. It just wasn't success in performance/$ in games without DLSS.

              Comment


              • #8
                Originally posted by rene View Post
                But, … nobody really into OpenSource uses Nvidia anyway, ... :-/
                Sadly OpenSource is not into ML and CUDA type workloads. They keep trying but cant quite catch up. ROCm is the best so far perhaps... but its going to take a couple of years to reach where CUDA is now.

                Comment


                • #9
                  Hey Michael, another great benchmark again.

                  BTW, I highly suggest you to include Redshift3D to your benchmarks. It has a commandline benchmark tool. Oh, it is CUDA only for now, but they just recently started to support AMD cards in OSX with Metal API and I bet that they will extend it on all OSes soon (obviously not with Metal).

                  Comment


                  • #10
                    Originally posted by Michael View Post

                    You can click on the OpenBenchmarking.org link, then hit the "normalize results" checkbox. You can also use 'highlight result' options if wanting to normalize against a particular GPU. And other options on OB to slice and dice the data however you want.

                    Hello Michael,

                    I feel you produce soo much information that I can't even keep up with it.
                    Have you thought of taking the data after the article(you have already done all the work) and then doing a youtube video going over it.
                    I was thinking maybe you would get money from advertisers on youtube or bring more traffic to this site?

                    Comment

                    Working...
                    X