Effective Runtime Fault Detection for DNN Accelerators

Systolic arrays are a popular choice for accelerating deep neural networks (DNNs) due to their inherent parallelism and efficient data reuse. However, ensuring the reliability of these DNN accelerators is crucial, as hardware faults can significantly degrade inferencing accuracy. Because systolic arrays utilize a large number of processing elements (PEs) for parallel processing of matrix multiplication, dataflow involving faulty PEs is especially of concern. Error propagation through PEs can reduce inferencing accuracy for DNN workloads. Even though many algorithm-based fault tolerance (ABFT) algorithms have been proposed to detect and correct errors in matrix multiplication, these ABFT methods cannot detect many errors originating from the accelerator hardware. We propose a run-time based fault detection technique leveraging functional data to generate checksums on-the-fly, avoiding the requirement for test patterns. Experimental evaluation shows that the proposed fault detection architecture can achieve 100% test coverage while incurring an area overhead of less than 2% for a 256 × 256 systolic array.