Updated GPU requirement for Einstein

Forums › Cruncher's Corner

AgentB
AgentB
Joined: 17 Mar 12
Posts: 915
Credit: 513211304
RAC: 0
18 Dec 2016 13:23:27 UTC
Topic 203850
(moderation:
    )

    With the new application requiring DP some older cards will stop, and DP performance can be very different relative to normal FP performance.

    There may be better lists but  https://en.wikipedia.org/wiki/List_of_AMD_graphics_processing_units and https://en.wikipedia.org/wiki/List_of_Nvidia_graphics_processing_units

    have some values for performance at DP.

    Milkyway have been running DP for some time.

    A quote from https://milkyway.cs.rpi.edu/milkyway/forum_thread.php?id=3188#57810

    NVIDIA:

    - Requires Compute Capability 1.3 and Above.
    - For the GeForce 2xx series, this is the GTX 260 and above.
    - Any Fermi based (GeForce GTX 4xx or 5xx) should support doubles.
    - Any older GPUs (such as a GeForce 8xxx or 9xxx) will not work

    AMD/ATI:
    - The oldest GPUs that work are the ATI HD Radeon 38x0 series.
    - In general laptop AMD GPUs do NOT support doubles despite similar branding (e.g. A Mobility Radeon 5870 is not the same as a normal Radeon 5870). Only the Mobility Radeon 48xx are the only current ATI laptop GPUs that have doubles.
    - For the Radeon 6000 series, only the 69xx have doubles.

     

    I'm nor sure what version of BOINC/ Drivers / OpenCL is needed (1.0 / 1.2 etc) for E@H, hopefully this thread can confirm

     

     

    Jim1348
    Jim1348
    Joined: 19 Jan 06
    Posts: 463
    Credit: 257957147
    RAC: 0

    I am not sure how much we can

    18 Dec 2016 15:32:12 UTC
    Message 153030
    (moderation:
      )

      I am not sure how much we can read into those comparisons, useful though they may be as a starting point.  For example, the GTX 960 has 2308/72.1 GFlOPS single/dual precision performance, while the HD 7870 has 2560/160 GFLOPS, and would seem to be the better card.  However, the times on my GTX 960s under Ubuntu have been in the range of 600-735 seconds, which are about as good as the times that Mad_Max reported for his HD 7870s (though they are on Win7)  https://einsteinathome.org/content/observations-fgrpb1-116-windows?page=2#comment-152999.

      So we will probably have to just report what we get on our cards, hopefully under comparable conditions (i.e., sufficient CPU support), and compare accordingly.  And I am hoping the AMD Vega cards bail us out of the DP performance shortfall.

      whore
      whore
      Joined: 5 Mar 15
      Posts: 7
      Credit: 8508986
      RAC: 0

      If I am reading this

      19 Dec 2016 2:47:59 UTC
      Message 153058
      (moderation:
        )

        If I am reading this correctly then my old GT 250 (1.1 capable) will no longer work for this project.  Is that correct?

        Mad_Max
        Mad_Max
        Joined: 2 Jan 10
        Posts: 154
        Credit: 2210161452
        RAC: 410475

        The double-precision

        19 Dec 2016 15:33:53 UTC
        Message 153085
        (moderation:
          )

          The double-precision requirements are not mandatory here. Only minor part of computations done in DP unlike Milkyway, main part of computation is still in single-precision.

          It is desirable to have DP card to get maximum speed , but tasks can run on GPU without DP support too. In this case DP part will be computed on main CPU. Look this message from Bernd Machenschalk:

          https://einsteinathome.org/content/gamma-ray-pulsar-binary-search-1-gpus?page=12#comment-152967

          So DP speed of card have only minor impact on computation time, leading factor is SP speed and VRAM speed (bus width, clock, cache efficiency).
          A big speed difference to be noticeable only for GPUs without DP support at all.

          AgentB
          AgentB
          Joined: 17 Mar 12
          Posts: 915
          Credit: 513211304
          RAC: 0

          Mad_Max wrote:The

          19 Dec 2016 22:29:36 UTC
          Message 153113 in response to message 153085
          (moderation:
            )

            Mad_Max wrote:
            The double-precision requirements are not mandatory here.

            Hmm, I wonder why we see a DP problem with HD 5830 https://einsteinathome.org/task/595454737

            From here

            Holmis
            Joined: 4 Jan 05
            Posts: 1118
            Credit: 1055935564
            RAC: 0

            There seems to be a bit of

            19 Dec 2016 22:58:31 UTC
            Message 153115 in response to message 153113
            (moderation:
              )

              There seems to be a bit of confusion about the requirement of DP, I too feel a bit confused. I found this message by Choks who helped develop the app in the Gamma-ray pulsar binary search #1 on GPUs thread, claiming that the app doesn't require DP.
              So maybe the error is caused by something else?

              whore
              whore
              Joined: 5 Mar 15
              Posts: 7
              Credit: 8508986
              RAC: 0

              Assuming my gpu (GT 250) is

              20 Dec 2016 0:08:26 UTC
              Message 153119
              (moderation:
                )

                Assuming my gpu (GT 250) is just no longer suitable for this project, then does anyone know of any other gpu projects which it could do?  The main reason I picked this project a year ago was because it still accepted my gpu and its paltry 1.1 compute capacity.

                The gpu works fine so I would hate to throw it away or recycle it if it can still do some good for someone.

                Richie
                Richie
                Joined: 7 Mar 14
                Posts: 656
                Credit: 1702989778
                RAC: 0

                whore wrote:The gpu works

                20 Dec 2016 1:26:25 UTC
                Message 153126 in response to message 153119
                (moderation:
                  )

                  whore wrote:
                  The gpu works fine so I would hate to throw it away or recycle it if it can still do some good for someone.

                  I understand that very well. I enjoy running little bit older electronics and like to exploit them till they sometimes quit cooperation. I don't know about your circumstances there. I looked at some comparisons and one thing I couldn't pass was the reality that GTS 250 has 150W TDP. That becomes more meaningful value when it's put in perspective with the computational speed. I only want to say maybe that somewhat high TDP is something to keep in mind. GTX 660 cards for example sell quite cheap used these days. Comparisons say that kind of 'next level up' GPU would have many times more speed for any project you would like to crunch... with lower power consumption at the same time. I think it's in the edge if GPU's from that generation (GTS 250) are rational to run for these kind of applications where it's supposed to run "pedal on the floor". Again, I understand well that kind of GPU still can do a nice job as a 'display adapter' in conventional daily usage.

                  AgentB
                  AgentB
                  Joined: 17 Mar 12
                  Posts: 915
                  Credit: 513211304
                  RAC: 0

                  OK just trying to get my two

                  20 Dec 2016 1:44:21 UTC
                  Message 153128
                  (moderation:
                    )

                    OK just trying (unsucessfully)  to get my two GTX-460s (768MB) to pick up tasks and scheduler logs reveal

                    [version] Checking plan class 'FGRPopencl-nvidia'
[version] parsed project prefs setting 'gpu_util_fgrp': 1.000000
[version] OpenCL GPU RAM required min: 1071644672.000000, supplied: 743047168

                    So that suggests you need a minimum of iGB RAM on the graphics card as well, so that might be another requirement.
                    Mad_Max
                    Mad_Max
                    Joined: 2 Jan 10
                    Posts: 154
                    Credit: 2210161452
                    RAC: 410475

                    AgentB wrote:Mad_Max

                    20 Dec 2016 3:05:26 UTC
                    Message 153131 in response to message 153113
                    (moderation:
                      )

                      AgentB wrote:
                      Mad_Max wrote:
                      The double-precision requirements are not mandatory here.

                      Hmm, I wonder why we see a DP problem with HD 5830 https://einsteinathome.org/task/595454737

                      From here

                      It is not a DP problem.
                      Yes, here is error about DP in logs, but this error is not critical - its intended: app try to use DP from GPU. But if it fail - it roll back to CPU for DP calculation. Here in log after error app writes:
                      "OpenCL device has no FP64 support" and continue normal work.

                      Real problem occur later:
                      ERROR: /home/bema/fermilat/src/bridge_fft_clfft.c:889: OpenCL memory allocation failed. status=-61

                      App can not get enough memory.

                      https://streamcomputing.eu/blog/2013-04-28/opencl-error-codes/

                      61 CL_INVALID_BUFFER_SIZE clCreateBuffer, clCreateSubBuffer if size is 0.Implementations may return CL_INVALID_BUFFER_SIZE if size is greater than the CL_DEVICE_MAX_MEM_ALLOC_SIZE value specified in the table of allowed values for param_name for clGetDeviceInfo for all devices in context.

                      I dont know why, but your card report what it can allocate 128 Mb of VRAM maximum:

                      Quote:
                      Using OpenCL device "Juniper" by: Advanced Micro Devices, Inc.
                      Max allocation limit: 134217728

                      And it is a real problem - app need about 600-800 Mb of VRAM to run.

                      Here log strings from my AMD GPUs with 1Gb and 2GB VRAM

                      Quote:

                      Using OpenCL device "Pitcairn" by: Advanced Micro Devices, Inc.
                      Max allocation limit: 804257792

                      Using OpenCL device "Pitcairn" by: Advanced Micro Devices, Inc.
                      Max allocation limit: 1877999616

                      Almost all VRAM available to task. In your case it looks like some sort of driver/openCL limitation.

                      Mad_Max
                      Mad_Max
                      Joined: 2 Jan 10
                      Posts: 154
                      Credit: 2210161452
                      RAC: 410475

                      AgentB wrote:OK just trying

                      20 Dec 2016 3:18:42 UTC
                      Message 153132 in response to message 153128
                      (moderation:
                        )

                        AgentB wrote:

                        OK just trying (unsucessfully)  to get my two GTX-460s (768MB) to pick up tasks and scheduler logs reveal

                        [version] Checking plan class 'FGRPopencl-nvidia'
[version] parsed project prefs setting 'gpu_util_fgrp': 1.000000
[version] OpenCL GPU RAM required min: 1071644672.000000, supplied: 743047168

                        So that suggests you need a minimum of iGB RAM on the graphics card as well, so that might be another requirement.

                         

                        Yes, in this case it is a real requirement unlike DP. App actually use 600-800 of GPU RAM while computing + some VRAM used by driver and OS, so admins limited work distribution to GPUs with at least 1 GB of RAM.

                        Comment viewing options

                        Select your preferred way to display the comments and click "Save settings" to activate your changes.

                        Forums › Cruncher's Corner