Rafael Tonetto, Pedro Pimenta, Abraham Chan, Karthik Pattabiraman, Fernando dos Santos, Marcello Traiola, Angeliki Kritikakou, Paolo Rech. To appear in the Proceedings of the International Conference for High Performance Computing, Networking, Storage, and Analysis (SC), 2026. (Acceptance Rate: 19.2%). [ PDF | Talk ]
Abstract: The increasing use of systolic arrays to accelerate AI in safety-critical domains, such as autonomous vehicles and space systems, makes their reliability a primary concern. Existing approaches often analyze the impact of faults at a single abstraction level, failing to capture how radiation-induced events at hardware level translate into system-level failures. This gap limits the ability to identify which hardware faults truly lead to critical application errors and to design effective mitigation strategies. To address this challenge, we propose a cross-layer analysis that traces faults from their origin in silicon to their impact on AI model outputs. By exposing tensor processing units to accelerated terrestrial and orbital radiation environments (neutron and proton beams) and combining hardware-level simulations with software-level fault injection, we identify specific error patterns and propagation mechanisms that result in mispredictions. This end-to-end perspective enables the identification of the low-level sources of errors that lead to critical failures.