A Robust DNN Accelerator with Data-path Fault Detection and Mitigation
TimeWednesday, December 8th6:00pm - 7:00pm PST
LocationLevel 2 - Lobby
DescriptionDeep Neural Networks are increasingly used in safety critical autonomous systems. We present a DNN accelerator architecture which provides fault detection and fault tolerance. Our approach is based on an Output Stationary systolic architecture augmented with fault detection based on fast on-line functional test of the Processing Elements (PEs). The hardware cost is under 8% and with Squeezenet, the loss in accuracy with a single faulty PE is under 3%, compared to 33% without mitigation. Dropout during training, without a priori knowledge of the faults, further improves fault tolerance.