GPU Parallel Computing

Jerry Shao, Ben Hsieh, and Song Yu
Supervisor: Zhangxing (John) Chen
84

Abstract

The purpose of this research is to use GPU computing power to optimize current serial sparse matrix calculation algorithms. Parallel Computing is that multiple executions are performed simultaneously. It is very important method to accelerate the calculation speed, and has been widely used to model scientific and engineering problems. The GPU is also very useful in the reservoir simulation area. GPU computing is great with Basic Linear Algebra Solver, especially with vector, dense matrices and sparse matrices calculation. High performance data parallelism on GPU can optimize data intensive calculation up to millions of variables. But there are some limitations in the GPU computing. The author`s research targets will be to optimize GMRES and BiCG algorithm for multiple GPU Computing on nVidia latest Tesla GPUs; optimize GMRES and BiCG algorithm for multiple CPU Computing on our IBM EXAS Cluster; and compare benchmarks from Single CPU, Multi-CPU and MultiGPU, target GPU speedups to 20x – 40x with thousands to millions of unknowns.