@techreport{oai:ipsj.ixsq.nii.ac.jp:00228897,
 author = {Shuming, Xu and Kazuteru, Namba and Shuming, Xu and Kazuteru, Namba},
 issue = {30},
 month = {Nov},
 note = {During the manufacturing phase of artificial intelligence (AI) chips, certain manufacturing faults bear profound significance due to their substantial impact on the precision of executed artificial intelligence wo rkloads. Detecting these critical functional faults necessitates the utilization of automatic test pattern generation (ATPG) tools, commonly employed to supply the requisite test patterns. However, such an approach entails notable costs and can potentially engender production losses. In this paper, we analyze the fault detection capacity predicated upon the criticality of chip functionality. Furthermore, we embark upon the task of categorizing faults within AI accelerators as either critical or benign. To this end, we present a machine learning architecture tailored to achieve fault detection in AI accelerators based on systolic arrays. Notably, we introduce an optimized generative adversarial neural network (GAN) WGAN-GP to ameliorate the misclassification challenges inherent to the designed detection architecture. Our results show that our method can identify faults with fixed accuracy and accurately classify them, thereby reducing production losses., During the manufacturing phase of artificial intelligence (AI) chips, certain manufacturing faults bear profound significance due to their substantial impact on the precision of executed artificial intelligence wo rkloads. Detecting these critical functional faults necessitates the utilization of automatic test pattern generation (ATPG) tools, commonly employed to supply the requisite test patterns. However, such an approach entails notable costs and can potentially engender production losses. In this paper, we analyze the fault detection capacity predicated upon the criticality of chip functionality. Furthermore, we embark upon the task of categorizing faults within AI accelerators as either critical or benign. To this end, we present a machine learning architecture tailored to achieve fault detection in AI accelerators based on systolic arrays. Notably, we introduce an optimized generative adversarial neural network (GAN) WGAN-GP to ameliorate the misclassification challenges inherent to the designed detection architecture. Our results show that our method can identify faults with fixed accuracy and accurately classify them, thereby reducing production losses.},
 title = {WGAN-GP based AI accelerator fault detection and fault classification analysis},
 year = {2023}
}