Simulation of read-disturb-induced inference accuracy degradation in an RRAM-based neuromorphic circuit

We establish a simulation based on a negative voltage read in a resistive random access memory (RRAM)-based neuromorphic circuit to investigate read-disturb-induced inference accuracy degradation. We characterize read-disturb behaviors in hafnium oxide RRAM cells at various read voltages and conductance levels. An analytical multi-conductance level read-disturb model is employed to describe a conductance change of an RRAM cell after repeated inference events. We simulate the read-disturb-induced inference degradation in image recognition of handwritten digits based on the Modified National Institute of Standards and Technology dataset. The impact of a read voltage and an RRAM conductance range for weight mapping on the inference accuracy degradation is evaluated.