The increasing demand for data-intensive analytics, driven by the rapid advances in artificial intelligence (AI), has led to the proposal of various AI accelerators. However, as AI-based solutions are being applied to applications that require high accuracy and reliability, ensuring the dependability of these solutions has become a critical issue. In this brief, we present an area-efficient systolic array redundancy architecture for reliable AI accelerator. In the proposed architecture, computations assigned to faulty multiply-accumulate (MAC) units are bypassed using dedicated routes. Subsequently, the same computations are executed in shiftable redundant MACs or selectable redundant MACs. This ensures the correct completion of calculations all without performance reduction. Moreover, the reassignment of computations can be efficiently managed through a simple scheduling algorithm. As a result, the proposed architecture achieves a high repair rate through the redundant MACs and effective computation reassignment. Despite these capabilities, the proposed architecture incurs only a small area overhead.
Research Foundation of Korea (NRF) grant funded by Korean Govern- In this brief, an area-efficient systolic array redundancy archi-ment(MSIT)underGrant2022R1A2B5B03002504.(Correspondingauthor: tecture is proposed. It bypasses faulty MAC units using ded-HayoungLeeiswiththeDepartmentofIntelligenceSemiconductorEngi-SunghoKang.) icated routes. Subsequently, computations associated with faulty neering,AjouUniversity,Suwon-si,Gyeonggi-do16499,SouthKorea(e-mail: MAC units are carried out using shiftable or selectable redun-hyleee@ajou.ac.kr). dant MACs. It ensures the correct completion of calculations all Jongho Park and Sungho Kang are with the Department of Electrical and without performance reduction. However, with only bypass log-ElectronicEngineering,YonseiUniversity,Seoul03722,SouthKorea(e-mail: ics inserted in each PE, the area overhead is highly reduced. Color versions of one or morefigures in this articlejongho0117@soc.yonsei.ac.kr;shkang@yonsei.ac.kr). In addition, a simple scheduling algorithm efficiently manages the https://doi.org/10.1109/TVLSI.2024.3421563. reassignment of computations and the control of input values for Digital Object Identifier 10.1109/TVLSI.2024.3421563 redundant MACs.This work was supported by the National Research Foundation of Korea (NRF) grant funded by Korean Government (MSIT) under Grant 2022R1A2B5B03002504.
