Resource-Aware Online Permanent Fault Detection Mechanism for Streaming Convolution Engine in Edge AI Accelerators

Abstract

Edge AI accelerators have gained popularity as a solution for applications such as image recognition sensors, remote sensing satellites, robotics, wearable devices, and drones due to their compact size and low power consumption. However, these applications demand fault tolerance, namely reliability, to overcome defects caused by radiation or manufacturing defects, especially in hard-to-reach environments like space or nuclear power stations. This paper presents an online permanent fault detection mechanism for streaming convolution engines in edge AI accelerators. The detection mechanism comprising extra comparison modules is added to a convolution engine’s processing elements (PEs). The experiment results show low overhead of the fault detection mechanism’s hardware resource and power consumption. The resource overheads are less than 3.6%, while the overhead of power consumption is not more than 1.2%.