Open-Source SpMV Multiplication Hardware Accelerator for FPGA-Based HPC Systems

The Sparse Matrix Vector (SpMV) multiplication kernel is a key component of many high-performance computing applications, but at the same time one of the most challenging to optimize, primarily due to its low flop-per-byte ratio and irregular memory accesses. As such, modern FPGAs, combined with High-Bandwidth Memory (HBM) modules, are much better-suited to the memory-bound nature of this kernel, compared to general purpose CPUs. Current FPGA-based approaches on SpMV support only single-precision floating point arithmetic. Moreover, they target for highly-streamed implementations that, although enhance performance, facilitate custom matrix storage formats, which (i) can increase the matrix footprint up to 3x, and (ii) drop the burden of input matrix transformation to developers. Towards widening the spectrum of FPGA-supported floating point formats for sparse algebra, this paper presents a first set of effective optimizations for double-precision SpMV hardware kernels using High-Level Synthesis (HLS) tools on HBM-equipped FPGAs. Results show that our work can provide 52.4x on average better performance compared to state-of-practice SpMV double-precision multiplication implementations on FPGAs for applications with volatile matrices, and up to 5.1x better performance-per-Watt compared to server-class CPUs.