Announcement

Collapse
No announcement yet.

Early 2016 Look: OpenCL & CUDA Comparison On 13 NVIDIA GPUs

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

  • Early 2016 Look: OpenCL & CUDA Comparison On 13 NVIDIA GPUs

    Phoronix: Early 2016 Look: OpenCL & CUDA Comparison On 13 NVIDIA GPUs

    With having out most of my NVIDIA graphics cards earlier this week due to running the 27-way OpenGL and performance-per-Watt comparison on NVIDIA graphics cards going back a decade, I took the opportunity to also run a smaller, fresh OpenCL/CUDA GPU compute comparison on various recent NVIDIA GPUs.

    Phoronix, Linux Hardware Reviews, Linux hardware benchmarks, Linux server benchmarks, Linux benchmarking, Desktop Linux, Linux performance, Open Source graphics, Linux How To, Ubuntu benchmarks, Ubuntu hardware, Phoronix Test Suite

  • #2
    The GTX 960 seems to make random appearances; it is only included in some of the charts.

    Comment


    • #3
      The 970 seems to be better in terms of performance per dollar compared to 960, but I'm going to save my money for Pascal. I have upgraded to a 450W PSU, yet my entire system with GTX 960 only drew about 180W max, which is not bad at all. I'm proud to have my graphics card for playing Bioshock Infinite, Metro: Last Light, The Witcher 2, and couple of other games that my AMD APU can't do well, which is sad since I would love to game with only an APU and not have a discrete dual-slot graphics card. I'm also proud to be able to use my GPU for use with Blender's Cycles engine for rendering 3D models.

      Originally posted by tpainter View Post
      The GTX 960 seems to make random appearances; it is only included in some of the charts.
      +1

      Comment


      • #4
        Seems to me that the 970 is the most well-rounded. Modestly good performance without being too expensive or power hungry.

        Comment


        • #5
          Seems like the 960 is really good if you are trying to do 2k gaming on the cheap.

          Comment


          • #6
            Not a single OpenCL vs CUDA benchmark ... Large 3D FFT, or large matrix multiplication are something challenging and available both from cuBlas/cuFFT and clBlas/clFFT.
            Would have been interesting, especially that clFFT is available for debian now.

            Comment


            • #7
              Originally posted by tpainter View Post
              The GTX 960 seems to make random appearances; it is only included in some of the charts.
              And 770 is also missing from the CUDA tests.
              Too bad, I searched for 970-770 comparion.
              I wonder everytime how nvidia could manage the double generation change, without decreasing the nanometres.

              Comment


              • #8
                What I don't understand is that why performing such test in single precision while OpenCL/CUDA are mainly designed for computing and so double precision is required? Strongly different results is expected in FP64 compare to what is shown here. Particularly, Maxwell series is weak in FP64 compared to Kepler and even Fermi because NVIDIA decided to cripple the FP64 units on the ship. I would recommend to perform real tests (i.e., FP64) is the current ones are note so useful. Just saying....

                Comment


                • #9
                  Originally posted by Goddard View Post
                  Seems like the 960 is really good if you are trying to do 2k gaming on the cheap.
                  It has always been marketed as such. Though idk if it's adequate enough when enabling high levels of AA and AF.

                  Originally posted by adakite View Post
                  What I don't understand is that why performing such test in single precision while OpenCL/CUDA are mainly designed for computing and so double precision is required? Strongly different results is expected in FP64 compare to what is shown here. Particularly, Maxwell series is weak in FP64 compared to Kepler and even Fermi because NVIDIA decided to cripple the FP64 units on the ship. I would recommend to perform real tests (i.e., FP64) is the current ones are note so useful. Just saying....
                  Taking your logic one step further, the test should have only included Titan cards, since those are the only ones really meant for OpenCL/CUDA.

                  Comment

                  Working...
                  X