A High-Performance Accelerator for Floating-Point Matrix Multiplication

Matrix multiplication is a widely-used routine in science and engineering applications. Accelerating this routine is important, because applications with large-scale matrix multiplication are increasingly common, especially in the area of high-performance computing (HPC). However, existing computing platforms including CPU, GPGPU and FPGA suffer from unsatisfactory performance or efficiency for this routine. In this paper, we propose a high-performance accelerator for double-precision floating-point matrix multiplication, and build a performance model for design space exploration based on a memory access scheduling. Impact of architecture parameters on accelerator performance and efficiency are evaluated and analyzed. Experimental results show that our proposed accelerator with 256 processing elements (PEs) can achieve a maximum performance of 767.99 GFLOPS and an efficiency of 99.99% for large-scale matrix multiplication, which is well suited to the requirement of HPC applications.