Towards Functionally Robust AI Accelerators

Recent advances in deep learning can be attributed to the continued performance improvement of hardware processors and artificial intelligence (AI) accelerators. In addition to conventional CMOS accelerators based on Von Neumann architecture, emerging technologies such as silicon photonics, memristors, and monolithic 3D (M3D) integration are being explored as post-Moore's law alternatives. However, the energy efficiency and performance of emerging AI accelerators can be catastrophically impacted by faults due to fabrication-process variations, thermal crosstalk, and aging. In this paper, we analyze the performance of several emerging AI accelerators in the presence of different uncertainties, and present low-cost methods to assess the significance of faults and mitigate their effects. We show that across all technologies, the impact of uncertainties on the performance can vary significantly based on the fault type and the parameters of the affected component. Therefore, the fault criticality-assessment techniques presented in this paper are necessary for yield improvement.