{"title":"Highly-Reliable Approximate Quadruple Modular Redundancy with Approximation-Aware Voting","authors":"Mahmoud Masadeh, Alain Aoun, O. Hasan, S. Tahar","doi":"10.1109/ICM50269.2020.9331771","DOIUrl":null,"url":null,"abstract":"Redundancy has been a general method to produce a fault-tolerance system. The Triple Modular Redundancy (TMR) with majority voters covers 100% single fault-masking, where the minimum area overhead is 200%. On the other hand, approximate computing is suitable for applications that can tolerate errors and imprecision in their underlying computations. Thus, inexact results allow reducing the computational complexity and hardware requirements with increased performance and power efficiency. This work explains how approximate computing could provide low-cost fault-tolerant architectures with an enhanced system’s reliability. In particular, we implement a novel Quadruple Modular Redundancy (QMR) designs using three identical approximate modules in addition to the exact module. Moreover, a two-steps magnitude-based voter is proposed to be able to tolerate approximation error. To validate our approach, we conducted experiments and the results showed the ability to achieve high fault tolerance, i.e., 99.88%, while reducing the probability of system failure by 15%, with 62% and 49.5% reduced area and power, respectively, compared to the traditional TMR.","PeriodicalId":243968,"journal":{"name":"2020 32nd International Conference on Microelectronics (ICM)","volume":"17 1","pages":"0"},"PeriodicalIF":0.0000,"publicationDate":"2020-12-14","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":"2","resultStr":null,"platform":"Semanticscholar","paperid":null,"PeriodicalName":"2020 32nd International Conference on Microelectronics (ICM)","FirstCategoryId":"1085","ListUrlMain":"https://doi.org/10.1109/ICM50269.2020.9331771","RegionNum":0,"RegionCategory":null,"ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":null,"EPubDate":"","PubModel":"","JCR":"","JCRName":"","Score":null,"Total":0}
引用次数: 2
Abstract
Redundancy has been a general method to produce a fault-tolerance system. The Triple Modular Redundancy (TMR) with majority voters covers 100% single fault-masking, where the minimum area overhead is 200%. On the other hand, approximate computing is suitable for applications that can tolerate errors and imprecision in their underlying computations. Thus, inexact results allow reducing the computational complexity and hardware requirements with increased performance and power efficiency. This work explains how approximate computing could provide low-cost fault-tolerant architectures with an enhanced system’s reliability. In particular, we implement a novel Quadruple Modular Redundancy (QMR) designs using three identical approximate modules in addition to the exact module. Moreover, a two-steps magnitude-based voter is proposed to be able to tolerate approximation error. To validate our approach, we conducted experiments and the results showed the ability to achieve high fault tolerance, i.e., 99.88%, while reducing the probability of system failure by 15%, with 62% and 49.5% reduced area and power, respectively, compared to the traditional TMR.