GPU cluster resources
Steve Brailsford
clusters after the AC cluster login expires? Is anyone offering a
subsription type service? Is there good ways my university can get
free/cheap hardware for academic research? I would also be interested
in a larger variety of configurations to test how certain changes
affect performance.
Thanks
Steve
Joshua A. Anderson
To get free hardware from NVIDIA, have your advisor sign up for the NVIDIA Professor Partnership program. It involves writing a small proposal discussing the science you are looking into and what hardware you need to get started. Note that NVIDIA will likely only give you the GPU, so you may need assistance from your IT support to setup or purchase a machine (or machines) to house what you get.
To get started, you really don't need a cluster and you probably already have what you need! Many laptops come with NVIDIA GPUs. While I cannot recommend running production simulations on laptops (you risk melting your computer), they are excellent test and development platforms. Certain generations of iMacs also have NVIDIA GPUs (you can check under Apple->About this Mac->More info->Graphics/Displays) the same overheating rule applies as iMacs are basically laptop hardware.
Similarly, many dell/hp/... desktops come with NVIDIA GPUs. What you have in yours might not be the fastest GPU - but still good for testing and development. And _any_ NVIDIA GPU less than 3.5 years old can run CUDA.
As for testing on a large variety of configurations, there is no easy solution someone out there can give you. In our lab, we have a test box I built from scratch - it has enough power and cooling to run 3 high-end GPUs. We've got a collection of them (some of the fastest of each hardware generation) which I benchmark on to make sure my code will run fast on all hardware. You can ask for these through the NVIDIA Partnership program - but desktop GPUs are cheap as far as research budgets are concerned, the absolute top of the line GTX 480 only costs $500. Go one step down to the excellent GTX 460 for only $200 - the 460 is easier to use in upgrading an existing system as it has less power and cooling demands than the 480. If you want to make the most of a request for free hardware from NVIDIA - ask for the Tesla C2050 or 2070.
If that route is not for you, you can always post a download link + compile instructions on the CUDA forums http://forums.nvidia.com/index.php?showforum=62 . The friendly regulars there will be happy to benchmark you code on whatever hardware they have, as long as you don't ask for it too often :)
--------
Joshua A. Anderson, Ph.D.
Chemical Engineering Department, University of Michigan