Efficient learning and crossbar operations with atomically-thin 2-D material compound synapses

Accurate and efficient synaptic weight programming and vector-matrix multiplication are demonstrated using compound synapses constructed with ultralow power binary memristive devices having oxidized atomically thin two-dimensional hexagonal boron nitride (BNO_x) filament formation layers. Experimental data of the resistive-switching current-voltage characteristics of BNO_x memristors are used to formulate variation-aware models that enable statistically analyzing the trade-off between efficiency and accuracy as a function of the synaptic resolution (i.e., levels of synaptic weight programming). Results are compared with commonly reported oxide-based memristors indicating orders of magnitude (i.e., ∼10⁵) improvements in power efficiency and ∼2-5× improvements in accuracy.