Invited Paper: A Holistic Fault Injection Platform for Neuromorphic Hardware

Logic-in-memory (LIM) is a promising flavor of the computing-in-memory (CIM) paradigm that utilizes memristive crossbar arrays to execute logic gates, resulting in high per-formance and energy efficiency. Binary neural networks (BNNs) can particularly benefit from LIM due to their massive parallel execution of binary logic gates. However, the impact of faults on BNNs accelerated with LIM has yet to be thoroughly investigated. To address this gap, we developed two distinct fault injection frameworks able to provide insights into the impact of different types of faults on the behavior of LIM. On the one hand, X-Fault aims to evaluate the impact of different faults that can affect crossbar arrays after manufacturing. On the other hand, FLIM allows for evaluating in-field faults on LIM. While both frameworks excel at their respective abstraction level, the complexity of neuromorphic systems requires a comprehensive fault analysis to grasp the fundamental impact stemming from the memristor to the BNN. Hence, we propose X-FLIM, a holistic fault injection platform capable of executing full-fledged BNNs on LIM while injecting in-field faults at the memristor and application level.