Nonideality Suppression and 16-State Multilevel Cell Storage Optimization in Phase Change Memory With Linear-Like Circuit

Abstract

Phase change memory (PCM) provides unique advantages in embedded, mass storage, and in-memory computing. Resistance drift and noise are decreasing the resistance ratio of PCM cell in amorphous and crystalline state, thereby reducing the readout window and data reliability, especially making obstacles to multilevel storage and neural network computing applications. A novel linear-like circuit is proposed to reduce drift coefficient and noise significantly and enhance the adjacent conductance ratio of the 16-level multilevel storage technology. Test results show that the drift coefficient reduced by 78%, and the noise reduced by three orders of magnitude. By constructing a PCM-based spiking recurrent neural network (SRNN) neural network in PyTorch, we achieved 97.4% accuracy, demonstrating that linear-like circuit could suppress drift and noise and improve the multilevel storage performance, which reduces the training loss despite PCM nonidealities.