Toward Functional Safety of Systolic Array-Based Deep Learning Hardware Accelerators

摘要: High accuracy and ever-increasing computing power have made deep neural networks (DNNs) the algorithm of choice for various machine learning, computer vision, image processing applications across spectrum. To this end, Google developed tensor unit (TPU) to accelerate computationally intensive matrix multiplication operation a DNN on its systolic array architecture. Faults manifested in datapath such due latent manufacturing defects or single-event effects may lead functional safety (FuSa) violation. Although DNNs are known resist minor perturbations with their inherent fault-tolerant characteristics, we show that classification model plummets from 97.4% 7.75% minimal fault rate 0.0003% accelerator, implying catastrophic circumstances when deployed mission-critical systems. Hence, ensure FuSa accelerators, article provides an extensive assessment accelerator exposed faults datapath, by varying network parameters, position, characteristics induced error multiple exhaustive data sets. Furthermore, propose two novel strategies obtain diminutive set test patterns detect violation accelerator. Our experimental results demonstrate obtained sets can achieve average 92.63% (in some cases, up 100%) coverage cardinality as low 0.1% entire set.