Approximate Early Output Asynchronous Adders Based on Dual-Rail Data Encoding and 4-Phase Return-to-Zero and Return-to-One Handshaking

TL;DR

This paper introduces robust asynchronous approximate adders using dual-rail encoding and 4-phase handshake protocols, demonstrating significant improvements in power, latency, and area over accurate asynchronous adders in CMOS technology.

Contribution

It presents the first implementation of approximate asynchronous adders based on delay-insensitive codes and 4-phase handshaking, with detailed analysis of their performance.

Findings

Approximate asynchronous adders reduce latency, power, and area.

Return-to-one handshake protocol yields greater improvements than return-to-zero.

Experimental results confirm the effectiveness of the proposed design in CMOS technology.

Abstract

Approximate computing is emerging as an alternative to accurate computing due to its potential for realizing digital circuits and systems with low power dissipation, less critical path delay, and less area occupancy for an acceptable trade-off in the accuracy of results. In the domain of computer arithmetic, several approximate adders and multipliers have been designed and their potential have been showcased versus accurate adders and multipliers for practical digital signal processing applications. Nevertheless, in the existing literature, almost all the approximate adders and multipliers reported correspond to the synchronous design method. In this work, we consider robust asynchronous i.e. quasi-delay-insensitive realizations of approximate adders by employing delay-insensitive codes for data representation and processing, and the 4-phase handshake protocols for data communication. The 4-phase handshake protocols used are the return-to-zero and the return-to-one protocols. Specifically, we consider the implementations of 32-bit approximate adders based on the return-to-zero and return-to-one handshake protocols by adopting the delay-insensitive dual-rail code for data encoding. We consider a range of approximations varying from 4-bits to 20-bits for the least significant positions of the accurate 32-bit asynchronous adder. The asynchronous adders correspond to early output (i.e. early reset) type, which are based on the well-known ripple carry adder architecture. The experimental results show that approximate asynchronous adders achieve reductions in the design metrics such as latency, cycle time, average power dissipation, and silicon area compared to the accurate asynchronous adders. Further, the reductions in the design metrics are greater for the return-to-one protocol compared to the return-to-zero protocol. The design metrics were estimated using a 32/28nm CMOS technology.