www.gpgpu.org

[rkane2]

Good afternoon,

I'm the system administrator for a general usage HPC cluster used for scientific computing. Running RH Linux 7.3.

Eight of the cluster nodes are visualization nodes with dual-head Matrox G400 graphics accelerators, which were considered fast a couple years ago. Many of the system users are doing calculations involving such things as Monte Carlo, FFT, sparse linear solvers, fluid dynamics, which are all things I understand to be good candidates for being passed off to the GPU.

What I am looking for is a quick way to evaluate whether using the Matrox cards is a reasonable way of getting some "free" gigaflops.

1) Does anyone have any experience with these cards.
2) Is there any software I can run without too much hassle that will give a reasonable expectation of what I can expect? (ie, to determine whether I spend considerably more time looking at the issue in depth)

Thanks,
Rob

[mhouston]

In short, no. The Matrox boards do not have floating point capabilites. For Linux, I would suggest upgrading to Nvidia 6800s. The Quadro versions, and maybe the higher end consumer versions, can support 4 outputs per board, which is probably why you purchased the Matrox boards in the first place. They will also GREATLY improve the performance of your vis cluster.

As for software, once you have the Nvidia boards, you can take a look at Brook for several benchmark style applications. Also, if I can ever find the time, we are hoping to finish the port of GPUBench to Linux.

-Mike

[mark]

I can't remember a time when those were considered fast. I guess they were the fastest (and only) 4-head beasts of their day.

Mark