Efficient and High-Performance Sparse Matrix-Vector Multiplication on a Many-Core Array

Abstract

Sparse matrix-vector multiplication (SpMV) is a critical operation in scientific computing, engineering, and other applications. Eight functionally-equivalent SpMV implementations are created for a fine-grained many-core platform with independent shared memory modules. These implementations are compared with a general-purpose processor (Intel Core-i7 3720QM) and a graphics processing unit (GPU, NVIDIA Quadro 620) and results are scaled to 32 nm CMOS. The performance (throughput per chip area) for all three platforms is compared when operating on a set of seven unstructured sparse matrices of varying dimensions up to 3.6 billion elements. The many-core implementations show a $54\times$ greater performance than the general-purpose processor, and $40\times$ greater performance than the GPU.