Highly Accurate and Energy Efficient Stochastic Multipliers

Abstract

Stochastic computing (SC) has gained interest by virtue of its small area and low power in computing units through encoding numerical values into randomly distributed bitstreams. However, this method not only results in long latency and thus increases energy consumption, but also reduces computing accuracy. Fortunately, stochastic bitstreams can be generated in parallel to reduce the latency. In this paper, parallel and serial multipliers using optimized encoding sequences with high accuracy and low energy consumption are proposed. Experimental results show that the mean squared error (MSE) is only 0.0015 for a 3-bit serial/parallel multiplier. The MSEs of 3-to 8-bit multipliers are reduced by approximately 29% on average compared with existing designs. The serial multiplier reduces the area and energy consumption by approximately 28% and 24% on average. The parallel multiplier reduces the latency and energy consumption by about 78% and 57% on average, respectively. To evaluate the effectiveness of the proposed multipliers, an image multiplication algorithm is implemented. The results show that the proposed multipliers yield higher image qualities than previous designs.