Runtime Fault Localization in Deep Neural Network 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, dataflow involving faulty PEs is especially of concern. Error propagation through PEs can reduce inferencing accuracy for DNN workloads. Although fault detection and repair techniques have been proposed to enhance the robustness of systolic arrays, fault localization remains an open problem. We propose a fault tolerance framework including run-time based fault detection and fault localization, both leveraging functional data to generate checksums on-the-fly. This approach enables error detection and localization during normal operation, avoiding the need for dedicated test patterns or additional downtime. Experimental evaluation shows that the proposed fault localization architecture incurs an area overhead less than 2% for a 256× 256 systolic array. In simulations, the proposed method achieves 100% fault detection and localization in a 256× 256 systolic array.