Aging-Aware Memristor Crossbar for Reliable and Energy- Efficient Deep Learning Acceleration

Abstract

Deep learning algorithms are highly energy and memory-intensive as their performance increases with an increasing amount of data. Moore’s law coming to an end and the ever-increasing demand for high computational power by Deep Learning algorithms are becoming a major issue. Another factor slowing down the fast Deep learning algorithms is the interconnect delay. This calls for a modern computing architecture that doesn’t physically separate memory and computation elements as done in Von Neumann's architecture. Memristor, the fourth fundamental circuit element, comes to the rescue. Owing to its less power consumption, more efficient and non-volatile nature, memristors claim to be a possible replacement for DRAM. Another advantage of memristor design is that it can be arranged in a crossbar arrangement. This makes it suitable to perform the dot-product operation and can be used in Convolutional Neural Network (CNN) architecture. BPhoton-CNN, proposed in this work, is a memristor-based CNN architecture that uses photonic Backpropagation for designing a complete analog system for training and inference. Despite showing the characteristics of a highly promising device for in-situ computing, memristive devices suffer from reliability issues given their non-linear nature. The proposed work also discusses the effect of one such non-linear characteristic called Aging. The effect of aging on the performance of deep learning accelerators and different methods to counter aging have been proposed in this work. heir engineering behavior. No significant correlation was observed among the other variables.