Dependable Neuromorphic Computing-in-Memory Architectures

Recently, neuromorphic computing has shown tremendous promise due to its energy efficiency and performance. Many academic and industrial initiatives focus on designing and developing neuromorphic platforms, either using classical complementary metal-oxide semiconductors or emerging exotic technologies. In particular, computing-in-memory (CiM) is one of the preferred paradigms for implementing these systems. Despite these efforts, ensuring the dependability of neuromorphic computing remains challenging due to various factors, such as device and technology non-idealities, aging electronics, fault tolerance, robust learning mechanisms, and security. These challenges hinder the adoption of emerging neuromorphic architectures in mainstream computing applications. To address this issue, this paper presents four consolidated works that tackle these challenges with innovative solutions, ranging from hardware-software co-design-based approaches to side-channel analysis and variability-aware device modeling and implementations.