A Comprehensive Analysis of Transient Errors on Systolic Arrays

Abstract

In recent years, the growth of interest in adopting deep neural network techniques across various domains led to new architectures for supporting the required computational effort. Tensor Processing Units (TPUs), which are based on a systolic array matrix multiplication unit (MMU), became widely popular thanks to their specific structure suitable for Artificial Intelligence. This work investigates Single Event Transient (SET) effects on TPU’s MMU. The analysis demonstrates the impact of SETs on the functionality of MMU when executing digital image processing filtering. The experimental results identify the static and dynamic SET sensitivity of TPU and depict meaningful information on the data dependency of the filters’ kernel values.