Ultra-short length stochastic computation based on multiple partition computing

In this paper, we propose an ultra-short length stochastic computation (SC) method based on multiple partition computing (MPC). For example, in MPC four parallel 16-bits streams can represent a 16-bits fixed-point number, while a 65536-bits stream is required in traditional SC. To achieve this, MPC partitions a fixed-point number to several short parts and converts each part to a short stream separately. A high performance stochastic multiplier and a divider are also proposed based on MPC. As a case study, we implement a 4×4 matrix multiplication with field programmable gate array (FPGA). The synthesis reports shows that the hardware efficiency of the proposed method is 76X times more than traditional SC and 23% improvement than the traditional fixed-point computation.