Efficient Parallel Stochastic Computing Multiply-Accumulate (MAC) Technique Using Pseudo-Sobol Bit-Streams

Abstract

Stochastic computing (SC) has emerged as a promising technique for reducing hardware costs in various applications, particularly in multiply-accumulate (MAC) intensive tasks such as neural networks. However, conventional SC still faces challenges in terms of achieving high accuracy and throughput. To enhance the precision, Sobol bit-stream has been widely adopted in SC. On the other hand, the throughput is frequently increased by means of parallel computing architecture. Nevertheless, directly increasing parallelism will incur significant additional hardware costs. In this paper, we propose Pseudo-Sobol bit-streams based on which an efficient parallel stochastic computing architecture for MAC operations is further developed. The proposed design leverages the properties of Pseudo-Sobol bit-streams and integrates the computation and conversion units to improve hardware efficiency. We evaluate the effectiveness of our design in two typical applications, general matrix multiplication (GEMM) and convolution. Experimental results show that our proposed design is capable of increasing energy efficiency by up to 36% and area efficiency by up to 70%.