Nonvolatile Resistive Memory Technology for Deep Neural Network Hardware Applications

Emerging resistive memory technology and a crossbar array platform play crucial roles in realizing the analog AI hardware. The memory devices store synaptic weights as conductance values, and the crossbar array physically implements a neural network (NN) on real hardware. Thanks to the compatible design with the NN structure, the resistive processing unit (RPU) system can accelerate NN inference and deep learning through parallel information processing in memory. However, realistic device behaviors must be considered to operate analog AI hardware successfully, such as conductance variability, retention, nonlinear conductance response to applied pulses, endurance upon several cycles of weight updates, etc. This chapter extensively explores the device technology for deep neural network (DNN) hardware implementation, including basic working principles of devices and arrays in the system, as well as requirements and challenges for implementing AI hardware from the literature. To support intuitive understanding and further exploration, this chapter also provides a MATLAB-based AI hardware simulator that implements analog in-memory inference and training of a HNN model, incorporating realistic device and array-level behaviors.