Utilizing ubiquitous commodity graphics hardware for scientific computing

Abstract

Current GPUs have many times the memory bandwidth and computing power compared to CPUs. The difference in performance is getting bigger as the evolution speed of the GPUs is higher than of the CPUs. This make it interesting to use the GPU for general purpose computing (GPGPU). I begin by looking at the architecture of the GPU, and some different techniques for programming on a GPU, including some of the available high-level languages. I have implemented the Mandelbrot computation on a cluster of GPUs (the HPDC display wall), and compared it against two different CPU implementations on the cluster. I have also implemented the Mandelbrot computation in both Cg and Brook, and compared the performance of the two languages. My experimental study shows that the GPU implementation of the Mandelbrot application is up to twice as fast as the load-balanced CPU implementation on the cluster of 28 computers, and up to 6 times faster on one computer.