Anyone using CUDA?
FedoraForum.org - Fedora Support Forums and Community
Results 1 to 3 of 3
  1. #1
    Join Date
    Jun 2004
    Location
    Edinburgh
    Posts
    641

    Anyone using CUDA?

    For a new project at work, I'm looking into building a processing farm of a few dozen Tesla 1U servers from Nvidia. This has lead us on to figure a way to let the developers run tests on their own workstations (with the appropriate gfx card).

    The cards we're looking at are Quadro FX 1800 or 3800's to provide quite a bit of poke.

    Thing is, I'm having issues with running any Cuda code successfully on systems installed with the pre-built nvidia drivers from RPMFusion. I can only get code running if I download the nvidia or cuda drivers from nvidia and install by hand. Not something I'd like to be maintaining across all out developer workstations.

    Is anyone out there running Cuda on Fedora/RHEL5 platform? If so, are you using RPM nvidia drivers or building them by hand?

    Thanks in advance

    Duncan

  2. #2
    Join Date
    Nov 2008
    Posts
    183
    I use CUDA on my system for some brute force work and other things. All I did in both a FC and Ubuntu system was install the drivers direct from Nvidia. Which is http://www.nvidia.com/object/linux_d...185.18.31.html for my 9x series cards on x64. I have had no issues. On Windows you have to download the CUDA code separate, or at least I had to last week for some stuff.

    FWIW this was on FC11 and Ubuntu 9.04

    To install the above driver, just download. Boot to runlevel 3 and as root run $sh nvidia_whatever_the_rest_is and follow the prompts.

  3. #3
    Join Date
    Jun 2004
    Location
    Maryland, US
    Posts
    7,406
    I'm using CUDA for two projects, one to accelerate Fast Fourier Transform calculations that some Java modelling code here at work used to do with a commercial math library. As more features were added the model runs were getting too long.
    The main modelling code is fine as is and profiling suggested that if only the FFT calculating could be done outside the Java with CUDA then the existing Java stuff could be 'saved' and get most of the GPU speed up. To do that I use JNI to adapt the Java to the CUDA (which is really like C) CUFFT library. JNI is an ugly way to call C from Java but it's all we have. CUDA 2.3 just came out and had two big improvements for us: double precision and some efficiency improvements for single precision.

    The other project is pretty similar but the code is in FORTRAN and again has to do with math similar to FFTs but this time it needs to compute huge numbers of 'sine transforms' (ST) which are related to FFTs. In this case the model has to reingest earlier ST calculations in a serial way so it's critical that we have double precision. Unfortunately CUDA's CUFFT library doesn't have sine transforms so we need to 'roll' our own sine transform. This can be done using a normal FFT CUDA call and some additionally code. The guy I work with is doing most of that since he knows Fortran.

    I don't install the rpmfusion drivers but use the run file from Nvidia. Our target machines have two Tesla C1060 cards, one model will use one Tesla card. with two models on each server.

Similar Threads

  1. nvidia cuda sdk examples won't compile [SOLVED]
    By jsnark in forum Using Fedora
    Replies: 6
    Last Post: 30th October 2009, 08:09 PM
  2. How to get Boinc 6.4.5 to use CUDA
    By Kobuck in forum Using Fedora
    Replies: 7
    Last Post: 15th September 2009, 05:50 PM
  3. Cuda on F11 x86_64
    By kramulous in forum Programming & Packaging
    Replies: 5
    Last Post: 20th July 2009, 03:18 AM
  4. Replies: 2
    Last Post: 15th November 2008, 06:16 AM
  5. NVIDIA CUDA & Livna
    By MarkS in forum Using Fedora
    Replies: 2
    Last Post: 24th June 2008, 06:33 PM

Tags for this Thread

Posting Permissions

  • You may not post new threads
  • You may not post replies
  • You may not post attachments
  • You may not edit your posts
  •