Announcement

Collapse
No announcement yet.

10-Way AMD & NVIDIA OpenCL GPU Linux Tests

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

  • 10-Way AMD & NVIDIA OpenCL GPU Linux Tests

    Phoronix: 10-Way AMD & NVIDIA OpenCL GPU Linux Tests

    For some weekend Linux benchmarking we tossed six NVIDIA GeForce graphics cards against four AMD Radeon graphics cards to get some idea for how the new OpenCL Linux benchmarks are running via the Phoronix Test Suite.

    http://www.phoronix.com/vr.php?view=19325

  • #2
    Well, that was a mixed bag for the binary drivers. Glad it isn't something I'm concerned about.

    Comment


    • #3
      Doesnt this show?

      Really, all these tests confirm is that despite OpenCL being 'vendor agnostic', you need vendor specific OpenCL code to get better results (or any) for different hardware?

      Some tests failing on AMD hardware completely, a few had certain nVidia GPU's not completing it, and the performance certainly didn't appear to scale with hardware for both vendors in a few tests. It's a mess really lol.

      Comment


      • #4
        lol @ nvidia beating AMD in openCL when nvidia has to manage two computing platforms(openCL and cuda)

        linux driver situation is still a mess.

        Comment


        • #5
          Originally posted by peppercats View Post
          lol @ nvidia beating AMD in openCL when nvidia has to manage two computing platforms(openCL and cuda)

          linux driver situation is still a mess.
          ...nvidia didn't win every test. As far as I'm aware, AMD joined pretty late with openCL on linux, so the fact that they managed to get ahead in any tests is actually pretty good.


          One thing people really don't seem to realize about GPGPU processing is even if you have something cross-platform like openCL, the process you're running is heavily dependent upon the hardware, NOT the drivers. Drivers make a difference but there's only so much they can do. A good example of this is bitcoin mining - it's an incredibly simple calculation, so AMD's overall higher frequency and higher stream processor count has a noticeable impact on bitcoin mining versus nvidia. There are $250 AMD GPUs that can beat out nvidia's Titan. However, that does not at all mean nvidia GPUs are bad pieces of hardware.

          You can say the exact same thing about Intel CPUs versus AMD. Some tests have shown that the FX series can even top Intel's 12-threaded i7s, but in most tests, the series fails because modern software isn't designed to operate the way AMD designed their CPU.

          Comment


          • #6
            Originally posted by schmidtbag View Post
            Some tests have shown that the FX series can even top Intel's 12-threaded i7s, but in most tests, the series fails because Intel pays companies to compile their software with ICC for amdcripple.
            fixed that for you

            Comment


            • #7
              I don't know what's more embarrassing, the test options or the fact those test options are 3 and 4 years old and designed primarily for CUDA and OpenMP?

              Comment


              • #8
                Originally posted by peppercats View Post
                lol @ nvidia beating AMD in openCL when nvidia has to manage two computing platforms(openCL and cuda)

                linux driver situation is still a mess.
                Well it is, but in Opencl it's at the same level as windows driver.

                It has problems with a lot of Opencl programs specially the ones with big kernels.

                A good example is render engines like Blender Cycles or Octane.

                Comment


                • #9
                  Originally posted by peppercats View Post
                  lol @ nvidia beating AMD in openCL when nvidia has to manage two computing platforms(openCL and cuda)

                  linux driver situation is still a mess.
                  It's just like peppercats said Nvidia has two computing platforms and has better OpenCL implementation than AMD by far.

                  But the funny fact is that this is not a linux driver issue, you can't run a lot of complex programs (probably depends on kernel size) with AMD Opencl in any OS, good examples are render engines like Blender cycles or octane.

                  With blender cycles and latest amd drivers linux is even faster than windows with cycles using Opencl but is still slower than a old core2quad CPU, and it only works in a few cards (GCN and some exceptions), with Nvidia it's working a lot faster with both Opencl and Cuda.

                  P.S. Did the 13.11 Beta 6 give some instability during the tests? i'm having some trouble with 3.12 kernel and this drivers... the open source ones works great but i have a 7950 so it's running too slow.

                  Comment


                  • #10
                    Sorry for the double post i thought i had hit some character limit in the first one... now i know that the post are moderated manually so it took time to be shown.

                    As i said in the reply above someone have problems with 13.11 b6? i have some crashes specially when using multiple monitors and when i disabled one through the AMD admin control center i have to do it again in the next reboot cause the setting is not saved :S

                    The Open source one works like a charm... but 7000 series runs pretty bad with 3d applications yet.

                    Comment


                    • #11
                      I think the tests should be run on windows to see if the driver is holding the cards back in Linux.

                      Comment


                      • #12
                        Originally posted by toyotabedzrock View Post
                        I think the tests should be run on windows to see if the driver is holding the cards back in Linux.
                        You will see simular results there. The more complex the computation, nVidia wins, the simpler the computation, AMD wins. Different architectures tuned to different tasks.

                        ( oh ya, you would see Titan results on openbenchmarking if Micheal ever figures out the login issue).

                        Comment


                        • #13
                          Originally posted by schmidtbag View Post
                          One thing people really don't seem to realize about GPGPU processing is even if you have something cross-platform like openCL, the process you're running is heavily dependent upon the hardware, NOT the drivers. Drivers make a difference but there's only so much they can do. A good example of this is bitcoin mining - it's an incredibly simple calculation, so AMD's overall higher frequency and higher stream processor count has a noticeable impact on bitcoin mining versus nvidia.
                          I'm a bit surprised by these results to be honest ; some years ago with pre Fermi card it was necessary to design gpgpu kernel around coalesced memory access, and it was said that developper was forced to use float4 type everywhere in their kernel to get good performance from evergreen/cayman radeon cards.
                          Nowadays Kepler and GCN cards uses cached memory and have a scalar ISA, which should make them much more "coherent" from a performance perspective except in some niche aera (for instance integer are much faster on radeon than geforce), but it's not really apparent in these benchmark.

                          Comment

                          Working...
                          X