AMCAL: Approximate Multiplier With the Configurable Accuracy Levels for Image Processing and Convolutional Neural Network

Multiplication is a necessary arithmetic operation in signal processing and machine learning algorithms. This paper proposes a novel method called Approximate Multiplier with Configurable Accuracy Levels (AMCAL) to obtain energy and area savings. The proposed algorithm performs the multiplication operation based on truncating the least significant bits and manipulating the remaining bits. Manipulation in the remaining input bits with a reduced bit width has been done so that the error resulting from truncating the bits is neutralized or reduced as much as possible. It has also shown that we are less concerned about the magnitude of the error injected in the input compared to other algorithms. The proposed AMCAL multiplier improves energy consumption by 86.7% compared to the Wallace multiplier, with an error margin of 0.15%. Moreover, in delay, power, and area, the proposed multiplier outperforms other approximate multipliers in the same class, such as DRUM and DSM. The AMCAL multiplier, surpasses new approximate multipliers such as DSI and TOSAM in power and area efficiency. Finally, it is shown that such a minor computational error does not affect the image quality resulting from four image processing programs, namely smoothing, sharpening, JPEG encoder and face alignment.