Graphics Processing Units (GPUs) were, as the name suggests, originally designed for the efficient processing of graphics. Over time, they were developed into systems that were capable of performing general purpose computing which is why some people refer to modern GPUs as GP-GPUs (General Purpose Graphical Processing Units).
Graphics processing tpically involves relatively simple computations that need to applied to millions of on-screen pixels in parallel. As such, GPUs tend to be very quick and efficient at computing certain types of parallel workloads.
The GPUComputing@Sheffield website aims to facilitate the use of GPU computing within University of Sheffield research by providing resources for training and purchasing of equipment as well as providing a network of GPU users and research projects within the University.
The University of Sheffield has been officially affiliated with NVIDIA since 2011 as an NVIDIA CUDA Research Centre. As such NVIDIA offer us some benefits as a research institution including discounts on hardware, technical liaisons, online training and priority seed hardware for new GPU architectures. For first access to hardware, training and details of upcoming events, discussions and help please join the GPUComputing google group.
Public GPU nodes have now been made available to Iceberg and ShARC users, these can be be used without acquiring extra permission.
Research groups also have an option to purchase and add nodes to the cluster to be managed by IT Services. For these nodes (e.g. Nvidia DGX-1 [COM]), permission from the group leader is required for access.
The node owner always have highest priority on the job queue but as a condition for the management of additional nodes on the cluster, the nodes are expected to be used as a resource for running short jobs during idle time. If you would like more information about having IT Services add and manage custom nodes, please contact research-it@sheffield.ac.uk.
Hardware, sofware and ways to access GPUs are different on Iceberg and ShARC. Instructions specific to each cluster can be found below: