Application-Induced Cell Reliability Variability-Aware Approximate Computing in TaOx-based ReRAM Data Center Storage for Machine Learning

Abstract

This paper proposes Variability-Aware Approximate Computing (V-AC) for TaOx ReRAM storage at data centers. For the first time, this paper shows that application-induced variability degrades the performance. To solve this problem, V-AC utilizes error resilience of machine learning (ML) application and reduces bit-error rate (BER) of typical cells by removing extra data copy and enlarging BER difference among cells. By combining device measurement and system emulations, this paper realizes system, circuit and device codesign (SCDCD). V-AC is key enabling technology to push the limits of performance, power, chip size and scaling of ReRAM for ML. Performance, energy and cell area of ReRAM storage improves by 7.0 times, 90% and 8.5%, respectively.