Transforming memristor noises into computational innovations

Abstract Memristor-based compute-in-memory (CIM) systems show promise in accelerating various computing tasks with high energy efficiency, while various inherent noises in memristors, generally viewed as non-ideal characteristics, are detrimental to system performances. However, recent studies reveal that these noises can be harnessed to enable advanced computational functionalities, transforming challenges into opportunities. In this work, we systematically review the noise utilization strategies for these functionalities by categorizing them into two main types: ‘noise-based perturbators’ and ‘noise-based generators’. The former utilize noise to help systems escape local minima and improve global convergence, as seen in combinatorial optimization, and to enrich feature spaces, as seen in reservoir computing (RC). The latter employ noise to produce random numbers or distributions, as used in physical unclonable functions (PUF), stochastic computing (SC) with true random number generator (TRNG), and Bayesian neural network (BNN). By examining these approaches, we highlight the potential of memristor noises to enable functionalities that are challenging to achieve with conventional precise computing systems. Finally, we discuss the challenges ahead and provide an outlook for future research. This review aims to pave the way for memristor-based energy-efficient and resilient computing technologies.