high performance computing on graphics processing units: hgpu.org

The paper introduces an algorithm which improves the value of the real giga floating point operations per second (GFLOPS) for matrix multiplication algorithm on Graphical Process Unit-GPU by overlapping the data transfers between (CPU) and the device (GPU) with the kernel execution. The input matrices are divided into n sections and the output matrix into n^2 sections. Streams are used to perform simultaneous data transfers and kernel executions in order to hide the memory copy operations. The results show that improved execution times and GFLOP values are obtained. The optimum value of n depends mainly on the matrix dimension and on the GPU type.