SPICE Simulation of RRAM-Based Cross-Point Arrays Using the Dynamic Memdiode Model

We thoroughly investigate the performance of Dynamic Memdiode Model (DMM) when used for simulating synaptic weights in large RRAM-based cross-point arrays (CPA) intended neuromorphic computing. The DMM is line with Prof. Chua’s memristive devices theory, which hysteresis phenomenon electroformed metal-insulator-metal structures represented by means two coupled equations: one equation current-voltage characteristic device based on an extension quantum point-contact (QPC) model dielectric breakdown and a second memory state, responsible keeping track previous history device. By considering ex-situ training CPA aimed at classifying handwritten characters MNIST database, we evaluate Write-Verify iterative scheme setting crosspoint conductances to their target values. total programming time, error, inference accuracy obtained such writing are investigated depth. role played parasitic components as resistance well some CPA’s particular features like dynamical range memdiodes discussed. interrelationship between frequency amplitude values write pulses explored detail. In addition, effect shift case programmed no errors studied variety input signals, providing design guideline selecting appropriate pulse’s frequency.