High Performance Matrix Multiplication on General Purpose Graphics Processing Units

In recent years, there has been significant interest from both academia and industry in applying commodity graphics processing units (GPUs) toward general computing problems. The nVidia CUDA programming model provides a straightforward means of describing inherently parallel computations. In this paper, we present our GPU-based matrix multiplication with high performance on General Purpose Graphics Processing Unit (GPGPUs). We implemented our algorithm using nVidia CUDA API and compared its performance with an optimized CPU-implementation on a high-end AMD Opteron Dual Core CPU. Our experimental results show that a significant performance improvement over CPU-based algorithm and the maximum observed speedups are about 100 times.