High-Precision Symmetric Weight Update of Memristor by Gate Voltage Ramping Method for Convolutional Neural Network Accelerator

Abstract

Memristor emerges as the key enabler for neural network accelerator. Here, we demonstrate high-precision symmetric weight update in a one transistor one resistor (1T1R) structure Ti/HfO <sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sub> /TiN memristor using a gate voltage ramping method, with over 120-level states and low variation (<; 4%). Incorporating all experimental non-idealities, the proposed mixed hardware-software convolutional neural network demonstrates over 92.79% online learning accuracy (against software equivalent 98.45%) for MNIST recognition task. The network also shows robustness to input image noises, array yield, and retention issues.