Announcement

Collapse
No announcement yet.

NVIDIA GeForce GTX 1660 Ti OpenCL Benchmarks, 14-Way NVIDIA/AMD GPU Compute Tests

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

  • NVIDIA GeForce GTX 1660 Ti OpenCL Benchmarks, 14-Way NVIDIA/AMD GPU Compute Tests

    Phoronix: NVIDIA GeForce GTX 1660 Ti OpenCL Benchmarks, 14-Way NVIDIA/AMD GPU Compute Tests

    On Monday we published the initial GeForce GTX 1660 Ti Linux benchmarks focused on gaming but due to having only a limited amount of time with that new Turing GPU at the time, CUDA/OpenCL benchmarks were yet to be completed. Our initial GPU compute tests with that "TU116" graphics card is now complete and we have those Ubuntu Linux benchmark results for sharing.

    http://www.phoronix.com/vr.php?view=27602

  • #2
    The geometric mean of the results seems to be completely dominated by the Turing advantage in the clpeak integer compute test. I wonder what this would look like without that one benchmark. GTX 1660 Ti behind the RX 590 and GTX 1070, likely?

    Comment


    • #3
      Originally posted by JanW View Post
      The geometric mean of the results seems to be completely dominated by the Turing advantage in the clpeak integer compute test. I wonder what this would look like without that one benchmark. GTX 1660 Ti behind the RX 590 and GTX 1070, likely?
      See the OpenBenchmarking.org result file linked in the article, there are also some other cl-mem tests where the 1660 Ti does well, etc.
      Michael Larabel
      http://www.michaellarabel.com/

      Comment


      • #4
        Thanks Michael this GPUComputing Bench,
        has today there are lots of people using graphics not to attach to a display( me included ), but for computing..

        So in floating point GPU computing AMD stack is in shape, but looses badly in Integer benchs?
        The latency of Rocm is still 2x the one from NV, but still Rocm 2.0 seems a good ground for a good GPU Computing stack..

        If team Red was able to cut this latency difference( that will take time since the driver now is very complex, with AMDGPU transition.. ), we will expect a big jump in performance from AMD.


        Also in the cl-mem:Copy bench, in absolute values is ok, but in Perf/Watt is lower,
        We already saw that in previous benchmarks, AMD needs to address that topic too..
        Last edited by tuxd3v; 03-01-2019, 11:39 AM.

        Comment


        • #5
          Originally posted by Michael View Post

          See the OpenBenchmarking.org result file linked in the article, there are also some other cl-mem tests where the 1660 Ti does well, etc.
          So I exported the results from OpenBenchmarking.org and looked at this. I don't quite get the exact same numbers as you for the geometric mean, since I don't know how to tell which results you include in there. I excluded all power consumption and temperature tests, and those not run on any AMD card. Anyway, my results seem close enough to yours.

          Basically, in the overall geometric mean relative to the GTX1660Ti, I find the GTX1070 to be 17% slower and the RX590 is 13-14% slower. Excluding just the "pts/clpeak-1.0.1 - --compute-integer" results changes this to the GTX1070 being 4% slower and the RX590 being 3% faster.

          When results are that heterogeneous, I guess summarizing them in a single number does not make much sense. My takeaway message would be "The GTX1660Ti is comparable in compute performance to the GTX1070 and RX590, except in OpenCL integer compute, where it is way ahead."

          Comment


          • #6
            and I just came back with the cheapest, useful AMD GPU I could find ;-) https://www.youtube.com/watch?v=f7C0INtRmfY hey, at least my cheap entry level card does not consume too many Watts and stays cool, ..! https://www.youtube.com/watch?v=wURWZpmgDV8
            Last edited by rene; 03-01-2019, 12:35 PM.

            Comment


            • #7
              Wow, this really reinforces that you have to select the right tool for the job, and ignore all the marketing hype. AMD dominates floating point performance with the ROCm OpenCL stack, meanwhile Nvidia wins on integer performance. And of course AMD wins on open source, while the Nouveau open source Nvidia driver remains unusable for any accelerated workload.

              Comment


              • #8
                Originally posted by JanW View Post
                My takeaway message would be "The GTX1660Ti is comparable in compute performance to the GTX1070 and RX590, except in OpenCL integer compute, where it is way ahead."
                Yes, I have the same feeling..
                In my opinion AMD played a safe bet, in Releasing the Rx590..

                Now NVidia has 2 cards with same performance GTX1660Ti, GTX1070 ..

                Comment


                • #9
                  Originally posted by rene View Post
                  and I just came back with the cheapest, useful AMD GPU I could find ;-) https://www.youtube.com/watch?v=f7C0INtRmfY hey, at least my cheap entry level card does not consume too many Watts and stays cool, ..! https://www.youtube.com/watch?v=wURWZpmgDV8
                  Woow nice setup
                  What is exactly the card you are showing in the 1st video at ~5:30?
                  A matrox graphics card?

                  Comment


                  • #10
                    Originally posted by tuxd3v View Post

                    Woow nice setup
                    What is exactly the card you are showing in the 1st video at ~5:30?
                    A matrox graphics card?
                    Yep, exactly. I have various vintage stuff from the last two decades, and for my YT content and hacking fun I even recently purchased one of the first ISA graphic accelerators (ET4000 w/32) and hacked some bare metal graphic acceleration code: https://www.youtube.com/watch?v=7cqj_aF80gw and a Vesa Local Bus accelerated, also because I never had one: https://www.youtube.com/watch?v=s1_LDJvOhmM if you are interested in vintage stuff, I also showed, used and tested various other vintage stuff, from a real am386, SparcStation, Sgi Octane, and stuff like that ;-)

                    Comment

                    Working...
                    X