Confined PCM-based Analog Synaptic Devices offering Low Resistance-drift and 1000 Programmable States for Deep Learning

摘要

We have demonstrated, for the first time, a combination of outstanding linearity of analog programming with matched PCM pairs, small analog programming noise, an extremely low resistance drift (R-drift) coefficient (0.005, median) and high endurance for a CVD-based confined phase change memory (PCM) with a thin metallic liner. In-depth analysis of linear analog programming is also presented. MNIST simulations using a pair of these confined PCM devices as a synaptic element yield a high test accuracy of 95%.