Fully-Parallel 2-Terminal Update Scheme for Tensor Product in ECRAM Arrays

We demonstrate fully-parallel tensor product in electrochemical random access memory (ECRAM) arrays enabled by multi-terminal operation to modulate ion injection into the metal-oxide channel. We show extensive characterization of the combined drain and gate pulsed response with a compact physics-based modeling to describe conductance update as a function of gate/drain voltages. We demonstrate fully-parallel product (AND) operation within the ECRAM array and accurate two-terminal weight update based on stochastic product, which paves the way for ECRAM-based hardware training accelerators. Finally, a realistic ECRAM model was calibrated on experimental data and used to simulate training of large convolutional neural networks (CNNs).