Toward Dependable AI/ML Hardware: Reliability Strategies Across Architectures and Memory Systems: Survey Paper

AI/ML workloads are becoming more widely used while their reliable hardware execution stands as a crucial factor especially when considering aggressive technology advancements and NVM technology adoption. The special session demonstrates extensive research into reliability enhancement methods for AI/ML hardware infrastructure at architectural and system-levels and design-time through neuromorphic and multiprocessor system studies. The discussion highlights three chief technical points which include elevated fault sensitivity of systems and aging mechanics due to voltage effects as well as the sustaining of reliability against hardware decay. The research shows breakthrough approaches which comprise Rox-ANN for efficient ANN implementation and PALP and DATACON methods for PCM memory optimization as well as MNEME and HEBE systems to manage NVM and hybrid memory aging and RENEU for neuromorphic reliability modelling under workload conditions. This work studies system-level methods that integrate thermal-aware task mapping and aging-aware checkpoint distribution and performs a combined analysis between DVFS and replication techniques in multiprocessor task scheduling. These solutions achieve substantial advancements regarding system energy efficiency coupled with increased performance while lengthening the system operational period. Experimental tests performed on SPEC CPU2017 and MiBench as well as diverse ML operations demonstrate the need for implementing early-stage reliability modelling approaches with workload management techniques that understand applications and architecture specifications to develop dependable AI/ML hardware systems.