Fault Tolerant Deep-Neural-Network Accelerators

In recent years, Deep Neural Networks (DNNs) have become increasingly popular for Artificial Intelligence (AI) applications because of their increase in accuracy compared to classical AI algorithms. They are, however, very computationally intensive and therefore specialized hardware, namely DNN accelerators, have been developed to be able to infer them efficiently. DNNs have also gained in popularity in harsh environment applications like image recognition in space satellites, autonomous vehicles, etc. However, currently no radiation hardened DNN accelerator exists. This project aims to solve this problem by developing a DNN accelerator and analyse how radiation introduces faults in this accelerator. Based upon this knowledge, efficient fault tolerance methods will be implemented using error detection based upon parity prediction. Finally, the design will be tested for fault tolerance after which a proof-of-concept ASIC will be created containing the radiation hardened accelerator.