From Emerging Memory to Novel Devices for Neuromorphic Systems: Consequences for the Reliability Requirements of Memristive Devices

摘要

Memristive devices have been developed initially for memories, where they function as a non-volatile memory element. While the different kinds of memristive devices have their own operational as well as reliability concerns, the specifications are clearly set by the memory application (e.g., embedded, standalone, or storage-class type). More recently, memristive devices became of high interest for enabling new hardware concepts for a variety of neuromorphic systems, ranging from machine learning, computation-in-memory, to full brain emulating systems. These applications are based on different functionalities of the memristive devices (as e.g. analog multi-level programming), and set different - but often not yet fully explored-reliability requirements. Interestingly, however, some of these neuromorphic circuits are more resilient to device failure, while major memory reliability threats as stochasticity, variability and noise even may become assets for building self-learning and predictive systems.