A Vector Systolic Accelerator for Multi-Precision Floating-Point High-Performance Computing

Abstract

There is an emerging need to design multi-precision floating-point (FP) accelerators for high-performance-computing (HPC) applications. The commonly-used methods are based on high-precision-split (HPS) and low-precision-combination (LPC) structures, which suffer from low hardware utilization ratio and various multiple clock-cycle processing periods. In this brief, a new multi-precision FP processing element (PE) is developed with proposed bit-partitioning method. Minimized redundant bits and operands are achieved. The proposed PE supports <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$16\times $ </tex-math></inline-formula> half-precision (FP16), <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$4\times $ </tex-math></inline-formula> single-precision (FP32) and <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$1\times $ </tex-math></inline-formula> double-precision (FP64) operations with 100% multiplication hardware utilization ratio. Besides, vector systolic structure is designed for PE array to increase the system-level throughput and energy efficiency. The proposed design is realized in a 28-nm process with 1.351-GHz clock frequency. Compared with the existing multi-precision FP methods, the proposed work exhibits the best energy-efficiency performance of 1193 GFLOPS/W at FP16, 317 GFLOPS/W at FP32 and 77.3 GFLOPS/W at FP64 with at least 22.3%, 30% and 3.3% improvement, respectively.