4
$\begingroup$

This is a follow-up question from linking the libraries for CUDALlink in Ubuntu 12.10.

The tests that I execute (CUDAQ[], CUDAInformation[], CUDADriver[]) seem to give successful outputs. However, when I try to execute actual CUDA commands the results are clearly wrong.

For instance, CUDATranspose[Table[i, {i, 2}, {j, 2}]] returns

    {{140191557896416,140191327433448},{17523916273012,0}}

And the numbers change each time I execute it. I also noticed that some commands complain that CUDALink is actually not initialised, like CUDADot[Table[i, {i, 2}, {j, 2}], Table[i, {i, 2}, {j, 2}]]:

    CUDADot::notinit: CUDALink is not initialized

But according to CUDAQ[], CUDAInformation[] and CUDADriver[] it should be. Before trying to solve the problem, how can I determine where it is? Is it within Mathematica? At the NVIDIA drivers level? Does it have to do with the gcc version?

Improve this question
$\endgroup$
4
  • $\begingroup$ Related to this ? $\endgroup$
    – b.gates.you.know.what
    Commented Apr 10, 2013 at 16:23
  • $\begingroup$ Very probably for the CUDADot part. Thanks $\endgroup$
    – Ziofil
    Commented Apr 10, 2013 at 16:26
  • $\begingroup$ Please come back with updates if you make any progress; there are a few people with the same issue. $\endgroup$
    – b.gates.you.know.what
    Commented Apr 10, 2013 at 16:27
  • 1
    $\begingroup$ I deactivated my linux licence and used it to test some CUDA commands on a windows 7 machine. It gives me very similar wrong results!! What is going on?? $\endgroup$
    – Ziofil
    Commented Apr 10, 2013 at 18:35

1 Answer 1

Reset to default
3
$\begingroup$

Changing graphics card to a newer one (GTX 660 Ti for me, from GeForce 9500 GT) worked for me. Probably there is an issue with using newer mathematica CUDA functions on older cards. I hope this helps.

Improve this answer
$\endgroup$
1
  • $\begingroup$ I think there is an issue with almost all CUDA functions on cards this old: they simply do not possess the hardware capabilities required by most useful workloads. The Wikipedia article does a good job of describing what is missing from CUDA 1.0. I'd also be surprised if the recent nVidia drivers had been properly tested with these old cards, so this problem may not be specific to Mathematica. $\endgroup$
    – Oleksandr R.
    Commented May 17, 2013 at 3:56

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.