This paper presents a novel analog resistive random-access memory (RRAM), named NeuroMem, which consists of Au/GO/Au. The device’s resistance can be tuned to any value within its R OFF to R ON range with high precision. The analog characteristic of NeuroMem mimics the memorization behavior of the brain, which makes it great asset for artificial neural network applications. In this work, NeuroMem-based crossbars are fabricated to hold the synaptic weights needed to perform Iris classification. The weight values are mapped to conductance states within NeuroMem R OFF to R ON range, and then written accurately on the actual devices. Unlike other RRAM-based hardware with limited conductance states, in this work no quantization is needed which enables efficient in memory-computing without scarifying accuracy. Furthermore, NeuroNem device has been demonstrated in crossbars on flexible polymer substrate using standard photolithography process, which facilitates producing low cost flexible electronics. This work opens up great insights towards realizing RRAM-based computing at the edge.
View full paper here: https://ieeexplore.ieee.org/document/9184484