Compact Modeling of Resistive Switching Memory (RRAM) With Voltage and Temperature Dependences

Resistive-switching memory (RRAM) devices are an attractive technology for in-memory computing (IMC) to accelerate data-intensive tasks, such as deep neural network (DNN) inference. However, the RRAM characteristics are significantly affected by voltage and temperature dependences, as well as by statistical variability of the programmed state, leading to accuracy degradations of the computed results. Compact models of RRAM need to account for these dependencies and variations to allow for accurate simulations of IMC circuits and design-technology co-optimization (DTCO). This work presents a compact model of RRAM devices that accounts for the conduction characteristics of various programmed states as a function of voltage and temperature. The compact model is validated through experimental measurements, and simulations of a simple use case of matrix-vector multiplication in a passive crosspoint array. The model can serve for the simulation and design of neural network accelerators based on RRAM technology.