Approximate multipliers based on a novel unbiased approximate 4-2 compressor

Abstract

Approximate computing, which relaxes full precision to achieve higher performance or lower power consumption, is widely used in error-resilient applications such as multimedia and machine learning. Multiplication is a fundamental operation for many of these applications. This study presents an unbiased approximate 4–2 compressor that generates positive and negative sign errors in balance. Based on the unbiased 4–2 compressor, two 8 × 8 unbiased approximate multipliers (UBAMs) are designed to meet various accuracy (or power) requirements. Experimental results indicate that one of the proposed designs has a smaller normalized mean error distance (NMED) than previous approximate multipliers, and the other offers a 39% smaller power-delay product (PDP) and almost 46% smaller energy-delay product (EDP). The proposed multipliers outperform other approximate designs in image filtering applications by achieving higher quality outputs with lower power consumption.