Optimizing The Performance Of A Cuda Kernel

Peter Pacheco, University of San Francisco

Stevenson Hall 1300
11:00 AM - 11:50 AM

General purpose programs for Graphics Processing Units (GPGPU) have become an extremely important tool in High-Performance Computing. However, because of the novelty of the architecture of GPU's, optimizing the performance of GPGPU's can be very challenging, and it may involve some subtle changes to the "obvious" code. We give a brief introduction to the CUDA API for programming Nvidia GPU's, and then discuss optimization techniques for a simple dot product.