上海科技大学知识管理系统

Unary computing is distinguished by its low hardware overhead and minimal power consumption, particularly excelling in multiplication tasks through bit-wise operations. However, conversions between binary and unary formats are commonly required in conventional unary computing systems, which may neutralize the hardware efficiency. In this work, we propose a unary multiplier that alleviates the cost of binary-to-unary converters by utilizing only one counter and one comparator for unary sequence generation. Leveraging a simple counting process, a counter with peripheral logic circuits is utilized to perform unary multiplication. Based on this technique, a general matrix multiplication (GEMM) unit, referred to as CUGEMM, is devised. To ensure a high accuracy in neural network computations, CUGEMM dynamically scales the small input values to avoid excessive accuracy loss. Experimental results reveal that the proposed CUGEMM achieves on average 14.71\% power reduction, while maintaining a high accuracy in ImageNet classification tasks compared to state-of-the-art (SoTA) designs.