Speaker
Description
In high radiation environments, hardware accelerators are prone to radiation-induced bit flips, leading to data corruption. For example, scientists at the Large Hadron Collider (LHC) seek to deploy hardware-accelerated neural networks in environments with radiation 1000X higher than that seen in space [1]. Ensuring reliable data collection, such as at the LHC, requires developing hardware that is resilient to data corruption, given the prevalence of such faults. Prior evaluations across available dataflows and data types were conducted to optimize power, performance, and area (PPA); however, evaluating dataflow and data type interplay for the resilience of common accelerator architectures remains underexplored [2][3]. Dataflow changes the order of operations and prioritization of data transfers across the memory hierarchy and compute datapaths. We develop a fault injection tool that accounts for quantization and dataflow. We evaluate the robustness of the dataflow and quantization under varying stuck-at-fault injection rates. Our work enables accelerator architects to build fault tolerance into the design.