A CGP-based Efficient Approximate Multiplier with Error Compensation

Abstract

As one of the most promising energy-efficient paradigms in deploying Neural Network (NN) on hardware, approximate computing ($A$ xC) has recently gained great traction to replace exact computing. This paper proposes an efficient approximate multiplier design method, which combines the Cartesian Genetic Programming (CGP)-based automatic design method and manual design method. Besides, an error compensation scheme based on the traversal search of truth table is proposed for higher-order multiplier construction. Experiments show that compared to exact multiplier, the proposed approximate multiplier can reduce the area, power consumption, and delay by 54.9%, 55.7%, and 36.86%, respectively. It also shows superiority to the state-of-the-art approximate multiplier. In addition, when deployed in LeNet-5 for MINIST datasets, the proposed multipliers show higher efficiency than exact multiplier with comparable recognition accuracy.