Radiation-Induced Soft Error Assessment of a Multithreaded MobileNet CNN Model Running in a Multicore Edge Processor

Abstract

Convolutional neural networks (CNNs) have become a standard technology in numerous industrial Internet of Things (IoT) applications and sectors, such as automotive and aerospace. Recent advancements in hardware and software (e.g., application programming interface (API)/libraries) components have enabled the efficient execution of multithreaded CNN models on edge devices. As the complexity and adoption of CNNs in safety-critical systems continue to grow, ensuring their resilience becomes key and increasingly challenging. In this context, this work promotes two original contributions: 1) the proposal of a multithreaded implementation of MobileNet, which achieves a $2.67\times $ speedup and an energy reduction of 16% with four worker threads, and 2) the first soft error reliability assessment of a multithreaded CNN model running in a multicore processor under high-energy and thermal neutron radiation flux. Results from the radiation campaigns, with more than 31k runs, suggest that multithreaded executions can increase the occurrence of critical faults by up to $5\times $ . Results also show a greater number of events during the thermal neutron campaign, and some input images are significantly more robust against silent data corruption (SDC) events.