PAKCK: Performance and power analysis of key computational kernels on CPUs and GPUs

Recent projections suggest that applications and architectures will need to attain 75 GFLOPS/W in order to support future DoD missions. Meeting this goal requires deeper understanding of kernel and application performance as a function of power and architecture. As part of the PAKCK study, a set of DoD application areas, including signal and image processing and big data/graph computation, were surveyed to identify performance critical kernels relevant to DoD missions. From that survey, we present the characterization of dense matrix-vector product, two dimensional FFTs, and sparse matrix-dense vector multiplication on the NVIDIA Fermi and Intel Sandy Bridge architectures. We describe the methodology that was developed for characterizing power usage and performance on these architectures and present power usage and performance per Watt for all three kernels. Our results indicate that 75 GFLOPS/W is a very challenging target for these kernels, especially for the sparse kernels, whose performance was orders of magnitude lower than dense kernels.