Announcement

Collapse
No announcement yet.

Running Some New Benchmarks On The GTX 1080 (Octane Render, OpenCL)

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

  • Running Some New Benchmarks On The GTX 1080 (Octane Render, OpenCL)

    Phoronix: Running Some New Benchmarks On The GTX 1080 (Octane Render, OpenCL)

    Several new benchmarks were added today to the Phoronix Test Suite / OpenBenchmarking.org and with their GPU-accelerated focus of the new tests, I couldn't help but run them on the mighty powerful GeForce GTX 1080 and friends...

    http://www.phoronix.com/scan.php?pag...New-Benchmarks

  • #2
    This mixbench repo has some results inside, unfortunately none of your tested cards match. What does the alternating and read-only modes mean? Which results could be comparable?
    Those also show performance loss starting with >64 iters for Nvidia's Maxwell and begnining with >128 for AMD, but only using OpenCL and CUDA, not for HIP. Which value did you take in your charts, the maximum result, the one with the highest iter count or some average value?
    Would be interesting to see some more HIP results
    Last edited by juno; 06-07-2016, 04:21 AM.

    Comment


    • #3
      Michael, do you still have a Titan Black (Kepler) ? I would love to see the FP64 comparison between 1080 and the TB (or the first release, i.e., "Titan") which was not crippled as opposed to the 780 Ti. Also, I'm surprised to see that in FP64, the 780 is better than 980. Is this really the case?

      thanks again for your great work!

      Comment


      • #4
        Michael, do you still have a Titan Black (Kepler)? This card is not crippled for FP64 and I'd love to see the comparison with 1080. I'm surprised to see better performances for the 780 compared to 980 in OpenCL FP64.

        Thanks for your work!

        Comment


        • #5
          Things are looking really good for GTX 1080. I am sure a subset of Phoronix readers would be interested in 1080's deep learning performance. There are indications that 1080 does not have double-rate FP16 performance (compared to FP32) but rather 1/64, so we'll be left with whatever advances were made in FP32 speed vs. Maxwell. Are you planning to test AlexNet anytime soon and post the results? A comparison to 980, 980Ti and Titan would be great.

          Comment


          • #6
            Originally posted by adakite View Post
            Michael, do you still have a Titan Black (Kepler)? This card is not crippled for FP64 and I'd love to see the comparison with 1080. I'm surprised to see better performances for the 780 compared to 980 in OpenCL FP64.

            Thanks for your work!
            I never had the Titan Black model. I had the original Titan but no longer have that.
            Michael Larabel
            http://www.michaellarabel.com/

            Comment


            • #7
              Originally posted by adakite View Post
              Also, I'm surprised to see that in FP64, the 780 is better than 980. Is this really the case?

              thanks again for your great work!
              That is, indeed the case. Being stuck at 28nm for so long, Nvidia increased Maxwell's performance by scaling back the double-precision power. A sensible trade-off for cards that are meant for gaming, but looks a bit weird in comparative tests like these.

              Comment


              • #8
                Originally posted by Michael View Post

                I never had the Titan Black model. I had the original Titan but no longer have that.

                All right, did you have any benchs in your archive with the original Titan so we can compare with this recent test of the 1080? Thanks again!

                Comment


                • #9
                  Originally posted by vocoloco View Post
                  There are indications that 1080 does not have double-rate FP16 performance
                  It's more like there are no indications that GP104 actually has this "mixed precision" ability - and in case it has, I'd expect Nvidia to restrict it for Tesla/Quadro cards.

                  Originally posted by adakite View Post
                  All right, did you have any benchs in your archive with the original Titan so we can compare with this recent test of the 1080? Thanks again!
                  The 780 Ti has a FP64/FP32 perf. ratio of 1/24, the original Titan 1/3. So you could multiply by 8 and deduce in which ballpark you might land ;-)
                  Maxwell and Pascal (GP104) have ratios of 1/32, which is why Kepler wins here, like bug77 said. However, it's quite impressive how the computing efficiency improved as the 780 Ti has more raw power than the 980 and still loses clearly in the integer/fp32 tests.
                  Last edited by juno; 06-07-2016, 09:03 AM.

                  Comment


                  • #10
                    Originally posted by adakite View Post


                    All right, did you have any benchs in your archive with the original Titan so we can compare with this recent test of the 1080? Thanks again!
                    Not with the same exact test profile versions.
                    Michael Larabel
                    http://www.michaellarabel.com/

                    Comment

                    Working...
                    X