Reliability Analysis of Approximate Multipliers with Recovery Schemes

Error tolerant applications such as video processing and machine learning can tolerate some kinds of errors in computation. For error tolerant applications, approximate computing, which aims to reduce computational cost and power consumption by omitting a part of computation within the error tolerance, has been attracting attention in recent years. As a kind of approximate computing with logic circuits for error tolerant applications, various approximate multipliers, which can produce approximate products with small hardware cost, have been proposed. Because such approximate multipliers are employed for error tolerant application requiring high reliability, e.g., autonomous systems, the reliability of the multipliers is one of important issues. In this paper, we focus on three types of approximate multipliers with recovery schemes and discuss the reliability of the approximate multipliers. According to fault acceptability, first, we propose a reliability model of approximate multipliers used for error tolerant applications. With this model, we analyze the relationship between the error recovery ability of the approximate multipliers and their reliability, and then discuss the design of approximate multipliers that can maximize their reliability. Experimental results clarify the optimal designs in terms of the reliability under several situations with different fault occurrence probabilities and error tolerance levels.