XNOR Operation of Binary Neural Networks Using Nanoelectromechanical Memory Switches

A highly efficient nanoelectromechanical memory switch-based binary neural network (NEM BNN) is proposed for the first time. Utilizing the electromechanical movement of a cantilever beam, XNOR operation for BNNs is implemented with two access transistors and an NEM memory switch. Owing to the unique properties of NEM memory switches with monolithic 3-D (M3D) integration and nonvolatility, the proposed NEM BNNs achieve 84% smaller area and 87% lower energy consumption than SRAM-based BNNs. Furthermore, owing to the superior ON/OFF resistance ratio of NEM memory switches, NEM BNNs feature higher energy efficiency, performance, and inference accuracy than other emerging nonvolatile-based BNNs.