Abstract
We are currently investigating the viability of many-core architectures for the acceleration of wavefront applications and this report focuses on graphics processing units (GPUs) in particular. To this end, we have implemented NASA’s LU benchmark – a real world production-grade application – on GPUs employing NVIDIA’s Compute Unified Device Architecture (CUDA).
This GPU implementation of the benchmark has been used to investigate the performance of a selection of GPUs, ranging from workstation-grade commodity GPUs to the HPC "Tesla” and "Fermi” GPUs. We have also compared the performance of the GPU solution at scale to that of traditional high perfor- mance computing (HPC) clusters based on a range of multi- core CPUs from a number of major vendors, including Intel (Nehalem), AMD (Opteron) and IBM (PowerPC).
In previous work we have developed a predictive “plug-and-play” performance model of this class of application running on such clusters, in which CPUs communicate via the Message Passing Interface (MPI). By extending this model to also capture the performance behaviour of GPUs, we are able to: (1) comment on the effects that architectural changes will have on the performance of single-GPU solutions, and (2) make projections regarding the performance of multi-GPU solutions at larger scale.