Presentation

DescriptionComponent reliability is a critical issue for deep learning accelerators based-on emerging technologies such as memristor crossbars (MC). Although MC-based deep learning is inherently tolerant to many components’ faults, aging is still a major concern that can lead to misclassification. In this paper, we present an aging-aware mapping approach for reliable deep learning in MCs. Results with a standard EDA for the CIFAR-10 data set and several benchmark DNNs demonstrate over 98.5% accuracy. In addition, the proposed approach reduces area-overhead up to 45% and improves energy savings by up to 15x compared to the state-of-the-art.