Large‐Scale and Highly Reliable Hopfield Neural Networks Using Vertical NAND Flash Memory for the In‐Memory Associative Computing

Large‐scale and highly reliable Hopfield neural networks (HNNs) for associative computing are implemented by vertical NAND (VNAND) flash memory for the first time. Fueled by the vertically stackable word‐line (WL) stacks and sequential page‐by‐page read‐out operations of a VNAND architecture, the HNN with 2 10 ‐nodes capable of asynchronous updates can be implemented in a single block, featuring remarkable proficiency in restoring 2 10 ‐bit (1 Kb) images corrupted by various types of noise. Utilizing asynchronous update scenarios, the VNAND HNN achieve a 13% increase in restoration accuracy and a 60.4% improvement in update speed compared with synchronous updates. Additionally, the influence of fabrication variation on accuracy degradation is analyzed. It is observed that the weights fine‐tuned by the incremental step pulse programming achieve <4% degradation rate. Furthermore, association performance is evaluated using restoration test from the untrained MNIST images, demonstrating a 98.7% association accuracy with the threshold ( θ ) modulation method for optimized asynchronous updates. Finally, thanks to the extremely low bit‐cost of a VNAND structure, the VNAND HNN occupies ≈117× smaller area than a 1T1R resistive random‐access memory HNN under the same technology node.