HPAM: An 8-bit High-Performance Approximate Multiplier Design for Error Resilient Applications
HPAM: An 8-bit High-Performance Approximate Multiplier Design for Error Resilient Applications
复制标题
HPAM：适用于容错应用的 8 位高性能近似乘法器设计
DOI：
10.1109/isqed54688.2022.9806220
复制
发表时间：
2022
期刊：
影响因子：
--
通讯作者：
D. Banerjee
中科院分区：
文献类型：
--
作者：
Divy Pandey;Vishesh Mishra;Saurabh Singh;Sagar Satapathy;Babita Jajodia;D. Banerjee
近年来，近似计算在功耗感知硬件架构设计中得到了广泛应用。近似计算技术可用于使一大类具有容错能力的应用受益。它已成为一种计算范式，能够有效地满足一些可以容忍结果存在有限不精确性的常见应用。图像处理、机器学习和深度学习等应用广泛使用对8位数字的乘法和加法运算。这项工作提出了一种用于容错应用的8位高性能近似乘法器（HPAM）。HPAM能够在应用端显著提高速度，同时保持较高的精度标准。其设计的动机是提供一个较宽的误差范围，从而使其在满足高精度需求以及低精度标准的应用方面都有价值。此外，与这项工作一起还提出了传统行波进位加法器（RCA）的一个近似版本，即分段行波进位近似加法器（SRCA）。为了验证所提出设计的有效性，将其性能与传统的华莱士树乘法器以及现有的最先进设计（如TOSAM、DSM和LETAM）进行了比较。平均而言，与现有的最先进设计相比，HPAM的速度提高了27.08%，结果精度提高了48.06%。
In recent times, approximate computing is widely employed in the design of power-aware hardware architectures. Approximate computing techniques can be used to benefit a major class of error-resilient applications. It has emerged as a computing paradigm that can efficiently cater several popular applications that can tolerate bounded imprecision in results. Applications such as image processing, machine learning, and deep learning extensively use multiplication and addition operations on 8-bit numbers. This work proposes an 8-bit High-Performance Approximate Multiplier (HPAM) for error resilient applications. HPAM is capable of providing significant speedup at application end while simultaneously maintaining high accuracy standards. It is designed the motivation of providing an broad error bound thus making it worthy in catering applications with high accuracy demands as well as low accuracy standards. Additionally, an approximate version of conventional ripple carry adder (RCA), a Segmented Ripple Carry Approximate Adder (SRCA) is also proposed along with this work. To validate the efficacy of the proposed design, its performance is compared with the conventional Wallace tree multiplier and the existing state-of-the-art designs such as TOSAM, DSM, and LETAM. On average, HPAM provides a speedup of 27.08% and 48.06% more accurate results in comparison to the existing state-of-the-art designs.