A Low-Power Approximate Multiply-Add Unit

Abstract

Modern applications such as image processing and deep learning are error tolerant in some level of inaccuracy. Approximate computing is one of the promising techniques that benefit applications in such domain by trading power for accuracy. This paper supposes applications that prioritize power and area over accuracy and proposes lower-power approximate multiply-add (MAC) unit with reduced area. The proposed MAC unit utilizes an approximate tree compressor (ATC), which was proposed in the previous study. From the experimental designs, it is unveiled that the proposed MAC unit consumes 48.7% less power and its area is 49.7% smaller than the conventional MAC unit.