Effect of OTS Selector Reliabilities on NVM Crossbar-based Neuromorphic Training.

This paper studies the use of 1-selector-1-resistor (1S1R) devices as the cross-point components in a crossbar array for deep neural network (DNN) training. Two training algorithms-minibatch gradient descent (MBGD) and a lowrank approximation to MBGD-are evaluated with compact models of ovonic threshold switching (OTS) selectors considering several reliability issues, and resistive randomaccess memory (ReRAM), a type of emerging non-volatile memory (NVM) device. We explore various reliability concerns of the OTS selectors including the low on-state conductance, threshold voltage variation, and internal voltage offset, etc. and propose corresponding techniques to overcome the reliability challenges.