A Fault Injection Framework for AI Hardware Accelerators

Abstract

Deep Neural Networks (DNNs) have proven to give very good results for many complex tasks and applications, such as object recognition in images/videos and natural language processing. Some relevant applications of DNNs are defined by real-time safety-critical systems, which typically require the adoption of DNN accelerators that are usually implemented as systolic arrays. Assessing their reliability is not trivial and may depend on several factors such as the size of the array and the data precision. In this paper, we present a cross-layer framework for systolic array DNN accelerators described at RTL level allowing to inject faults at channel granularity for convolutional layers. The basic idea is to simulate the execution of the Channel Under Test (ChUT) at RTL level. Faulty outputs collected from the RTL simulation are then used at software level to complete the execution of the DNN and thus determine the impact of the injected faults at application level. Interestingly, the software execution is more than 100 times faster than the corresponding hardware simulation.