Design and Analysis of Energy Efficient Approximate Multipliers for Image Processing and Deep Neural Network

Numerous obstacles in enhancing the performance of computing systems have spurred the emergence of approximate computing. Extensive studies have been reported on approximate computing to develop high-performance, energy-efficient hardware designs tailored to error-resilient applications. In this brief, we proposed 8-bit approximate multipliers with 15 levels of accuracy using three techniques: recursive, bit-wise, and hybrid approximation using partial bit OR (PBO). Compared to the existing multipliers, investigated designs have significantly improved the area, power, delay, Power Delay Product (PDP), and Power Area Delay Product (PADP) by 41.68%, 73.16%, 35.57%, 72.65%, and 75.42% respectively on average. On resemblance with the accurate multiplier, the area, power, delay, PDP, and PADP were enhanced by 54.41%, 57.57%, 25.73%, 60.14%, and 74.33% correspondingly on average. Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index Measure (SSIM) values surpassing (30 dB, 94%), (31 dB, 96%), and (26 dB, 95%) by applying them to benchmarks in image smoothing, edge detection, and image sharpening successively. Moreover, upon scrutinizing the efficacy of multipliers in hardware implementations of deep neural networks attaining the performance exceeding 95%. The obtained results confirm that suggested multipliers are well-suited for their widespread applications.