Overview on using GPUs

GPUs provide the opportunity to do massively parallel computation, in particular executing the same computations on many inputs. To do this, you either need access to a machine with an installed GPU and appropriate software libraries, or you need to use a virtual machine in the cloud.

GPU Resources

• The SCF has a GPU installed on one of the nodes (roo) of our cluster as well as a large number of other GPUs purchased by or donated to specific faculty members that are also available for use on a preemptible basis.
• The campus-wide Savio cluster has GPU nodes.
• GPUs are available through NSF's ACCESS network of supercomputers and on commercial cloud providers such as Amazon's AWS and Google cloud. Email us for more information.

For more details on the SCF GPUs and on using GPUs on Savio, please see this link.

For the most part, researchers tend not to program directly on a GPU but to use libraries such as PyTorch and JAX for Python (also CUDA.jl for Julia) that make use of GPU(s) without one having to specifically write GPU code. However, this tutorial provides some information about programming directly on a GPU.