Subject: | |
From: | |
Reply To: | |
Date: | Fri, 10 Aug 2007 23:53:38 +0100 |
Content-Type: | text/plain |
Parts/Attachments: |
|
|
On Fri, 2007-08-10 at 13:22 -0500, Nathan Moore wrote:
> Hi,
>
> I recently installed an NVIDIA 8600GT graphics card (PNY brand) in an
> AMD box running SL5, x86/32bit version. Rather than using for
> graphics, I plan to use the card as a linear algebra co-processor ( a
> fun project for students in computational physics...). NVIDIA
> provides the CUDA environment/library
> ( http://developer.nvidia.com/object/cuda.html) to facilitate using a
> graphics card for numerical work.
>
Nathan, I bought an 8600GTX card a couple of weeks ago, with similar
intentions to evaluate CUDA before the launch of the dedicated Tesla
cards (and I doubt I'll ever have a Tesla card at home to play with
anyway!)
I run SuSE 10.2 at home (on a Sun Ultra 20 workstation). The CUDA
toolkit and SDK installed without problems, and the examples run fine.
libtlshook.so is part of the CUDA install - I installed under /usr/local
so it is /usr/local/cuda/lib/libtlshook.so
You need to set LD_LIBRARY_PATH such that $CUDA_ROOT/lib is included
(substitute $CUDA_ROOT for where you installed CUDA)
Cue me now boring the list by saying that the last time I worked with
vector processors was on an IBM 3090. I compiled up part of the Aleph
Monte Carlo package on it, and it ran slower. Admittedly this was a very
naive attempt - just throw it at the compiler and hope for the best.
|
|
|