A Behavioral Model of Digital Resistive Switching for Systems Level DNN Acceleration

The deployment of IoT has brought on the generation of massive amounts of data in need of analysis. In recent times, resistive switching-based crossbar arrays have been presented as a viable candidate for the acceleration of neural network inference, in pushing beyond the limit of CMOS process scaling so as to keep pace with the ever-growing complexity of computation. While plenty of empirical and physically descriptive models exist, their simulation run times become inconvenient for users when used in large scale crossbar arrays. In this brief, we first present a behavioral model of digital resistive switching devices, demonstrated on experimental PCM data to exhibit generality which can see useful implementation in circuit analysis methods for compute-in-memory applications. This model is based on a pair of nonlinear ordinary differential equations that request switching time and threshold voltage inputs from the user, which are the most important concerns for binarized weights in crossbar arrays. By stripping the model of detailed physical characteristics that is not required at the systems level, we demonstrate an improvement of computational run time of up to 20-fold over state-of-the-art physics-based models, and 1.3 times over the most commonly used empirically driven models.