GPU Compute Power


Advanced search

Message boards : Number crunching : GPU Compute Power

AuthorMessage
anthonyshea
Send message
Joined: Nov 14 12
Posts: 2
Credit: 96,491
RAC: 0
Message 210 - Posted 24 Jan 2013 8:55:05 UTC

    have you thought about rewriting your code to take advantage of the massive amount of GPU processing power available in the BOINC network?

    World Community Grid (IBM sponsored BOINC project)has done this and my work units have gone from a few hundred a day to a couple of thousand!and I have a fairly out dated GPU (ATI HD5850).

    Dagorath
    Avatar
    Send message
    Joined: Oct 15 11
    Posts: 11
    Credit: 63,227
    RAC: 0
    Message 211 - Posted 24 Jan 2013 12:39:23 UTC

      anthonyshea is correct. BOINCers are buying GPUs and putting them to work on projects. The numbers grow every day! Look at this list of hosts of one GPUgrid volunteer. You see 5 top of the line desktop GPUs plus 2 medium power GPUs in his list. And he is only number 3 on the Top Volunteers list!!

      I will not buy any more expensive powerful CPUs like i7 because they cannot do even 10% of the work a GPU can do. They are a waste of money. From now on I will buy only cheap CPUs like i3 or even less, just enough CPU to drive a powerful GPU.

      If you don't grow and improve the project to keep up with new technology you will be left behind.

      anthonyshea
      Send message
      Joined: Nov 14 12
      Posts: 2
      Credit: 96,491
      RAC: 0
      Message 212 - Posted 26 Jan 2013 9:39:53 UTC - in response to Message 211.

        Last modified: 26 Jan 2013 9:42:27 UTC

        Yeah, I plan on a i7 build but on the 2011 socket,because I want the quad memory controller and a hexicore cpu. because of other creative programs I use, but I'm by far not the average consumer. and an i5 is more than capable of running the most demanding games out on the market and will continue to be so for quiet a while to come.on the other hand people are constantly upgrading their gpu's. it just seemed like a natural transition to me, on a workload point of view for them to make. the future will tell.

        Profile Oleg Zaikin [SAT@home]
        Forum moderator
        Project administrator
        Project developer
        Project scientist
        Send message
        Joined: Sep 15 11
        Posts: 133
        Credit: 4,826,453
        RAC: 0
        Message 215 - Posted 7 Feb 2013 15:07:11 UTC - in response to Message 210.

          Last modified: 7 Feb 2013 15:07:34 UTC

          have you thought about rewriting your code to take advantage of the massive amount of GPU processing power available in the BOINC network?

          World Community Grid (IBM sponsored BOINC project)has done this and my work units have gone from a few hundred a day to a couple of thousand!and I have a fairly out dated GPU (ATI HD5850).


          You (and Dagorath) are right, GPU is very major for a project. We are working on it, see new section about testing of GPU on the site. I hope we will launch it on this year.

          Dagorath
          Avatar
          Send message
          Joined: Oct 15 11
          Posts: 11
          Credit: 63,227
          RAC: 0
          Message 220 - Posted 8 Feb 2013 6:04:38 UTC - in response to Message 215.

            Excellent! I will download and run it in 2 minutes. If it's stable and well behaved I will announce it on the BOINC dev forums and at GPUgrid where there are many nVIDIA systems willing to help, if you approve.

            Dagorath
            Avatar
            Send message
            Joined: Oct 15 11
            Posts: 11
            Credit: 63,227
            RAC: 0
            Message 221 - Posted 8 Feb 2013 7:44:06 UTC

              I downloaded both files and saved the text file as in.txt. When I ran the binary it said I am missing libcudart.so.4. I have libcudart.so.4.2.9 which is not exactly the same but I copied it to /usr/lib/x86_64-linux-gnu/libcudart.so.4.

              I was also missing libboinc_api.so.7 so I installed package boinc-dev from Ubuntu 12.10 Software Centre.

              Then the binary ran, for about 5 seconds. Output to the terminal was:

              kim@norman:~/sat_gpu_test$ ./ringsat-gpu20_ls92

              Complete tasks: 0
              UC size: 10 kim@norman:~/sat_gpu_test$

              It could use and extra CR-LF at the end. More important... it created out.txt but it's empty.... 0 bytes.

              It created boinc_lockfile, out.txt and stderr.txt. Stderr.txt says:

              kim@norman:~/sat_gpu_test$ cat stderr.txt
              Hello, stderr!
              00:00:07 (353): Can't open init data file - running in standalone mode
              cudaCheckError() with sync failed at ./ringsat_gpu.cu:1055 : the launch timed out and was terminated.
              kim@norman:~/sat_gpu_test$

              Profile Oleg Zaikin [SAT@home]
              Forum moderator
              Project administrator
              Project developer
              Project scientist
              Send message
              Joined: Sep 15 11
              Posts: 133
              Credit: 4,826,453
              RAC: 0
              Message 224 - Posted 9 Feb 2013 12:24:53 UTC - in response to Message 220.

                Excellent! I will download and run it in 2 minutes. If it's stable and well behaved I will announce it on the BOINC dev forums and at GPUgrid where there are many nVIDIA systems willing to help, if you approve.



                It is not stable, thanks for error report. I've sent your message to our GPU programmer.

                vader
                Send message
                Joined: Sep 24 11
                Posts: 1
                Credit: 679,734
                RAC: 0
                Message 225 - Posted 9 Feb 2013 16:17:22 UTC - in response to Message 221.

                  Hello! I'm the developer of the GPU app for the SAT@home project.
                  The "timeout" bug you just experienced is not a bug?, it's a watchdog feature of the nvidia driver that kills the GPU app after some short time, if it runs on a main (X server screen) GPU without responding to system calls.
                  If you want to test our app you should disable CUDA watchdog through environment variable or test in on a separate, no-X-server GPU.

                  Profile Coleslaw
                  Send message
                  Joined: Oct 18 11
                  Posts: 3
                  Credit: 1,032,653
                  RAC: 0
                  Message 284 - Posted 19 Nov 2013 23:46:39 UTC

                    So, it has been 9 months... is there an update on GPU capabilities?
                    ____________

                    ZloyZayac
                    Send message
                    Joined: May 8 13
                    Posts: 3
                    Credit: 45,843
                    RAC: 0
                    Message 287 - Posted 7 Jan 2014 8:31:05 UTC

                      Why you do not declare the collection of donations for the development of OpenCL applications?

                      How much will it cost if hiring a freelancer?

                      sergei.a.mochalov
                      Send message
                      Joined: Jul 8 13
                      Posts: 7
                      Credit: 1,205,665
                      RAC: 0
                      Message 288 - Posted 16 Jan 2014 22:45:48 UTC - in response to Message 287.

                        I think I've seen these ideas before... I think that I see the hare's long ears... :)

                        Mark Sechin
                        Send message
                        Joined: Jun 15 14
                        Posts: 3
                        Credit: 1,315,908
                        RAC: 0
                        Message 811 - Posted 5 Feb 2015 23:50:02 UTC

                          One year passed. Are there any updates for GPU utilization?

                          Post to thread

                          Message boards : Number crunching : GPU Compute Power


                          Home | My Account | Message Boards


                          Copyright © 2019 Institute for System Dynamics and Control Theory of SB RAS and Institute for Information Transmission Problems of RAS