FSA: An Efficient Fault-tolerant Systolic Array-based DNN Accelerator Architecture
TimeTuesday, July 12th6pm - 7pm PDT
LocationLevel 2 Lobby
Event Type
Networking Reception
Work-in-Progress Poster
DescriptionPermanent faults have been a critical problem for Deep Neural Network (DNN) accelerators, as they lead to undesirable ramifications that can severely impact DNN inference accuracy. In this paper, we propose FSA, a novel fault-tolerant systolic array-based DNN accelerator design with the purpose of maintaining DNN inference accuracy without incurring significant computational latency and power consumption in the presence of permanent faults. The key feature of the proposed FSA is a unified re-computing module (RCM) to recalculate the required DNN calculations mapped to faulty processing elements (PEs) of the DNN accelerator regardless of the number and location of faulty PEs.