HyCA: A Hybrid Computing Architecture for Fault-Tolerant Deep Learning

Hardware faults on the regular 2-D computing array of a typical deep learning accelerator (DLA) can lead to dramatic prediction accuracy loss. Prior redundancy design approaches typically have each homogeneous redundant processing element (PE) mitigate faulty PEs for limited region rather than entire avoid excessive hardware overhead. However, they fail recover when number in any exceeds same region. The mismatch problem deteriorates fault injection rate rises and are unevenly distributed. To address problem, we propose hybrid architecture (HyCA) fault-tolerant DLAs. It has set dot-production units (DPPUs) recompute all operations that mapped despite PE locations. According our experiments, HyCA shows significantly higher reliability, scalability, performance with less chip area penalty compared conventional approaches. Moreover, by taking advantage flexible recomputing, also be utilized scan detect effectively at runtime.