Hardware-Efficient Nonlinear Voltage Threshold Estimation for NAND Flash Memory via Lightweight Deep Learning

Modern NAND flash memory requires adaptive signal processing to address reliability degradation caused by technology scaling and environmental stress. Conventional NAND flash controllers rely on static, block-level read voltage thresholds (VTs), which fail to capture row-level variations and lead to increased read-retry latency. This work proposes a real-time, nonlinear VT estimation framework for high-throughput NAND flash controllers. The framework enables per-row VT adaptation, while incurring minimal metadata overhead. In particular, VTs are estimated using lightweight deep learning techniques, including deep neural networks (DNNs) with fully-connected hidden layers and entity embeddings. For read-retry scenarios, we introduce a DNN-based VT estimator that leverages optimized sparse sampling of the VT distribution using a small number of additional reads. A hardware-efficient read flow integrating the proposed algorithms is presented. Experimental results on Quad-Level Cell NAND flash memory devices demonstrate significant reductions in added bit error rate compared to fixed and linear VT methods, while maintaining peak performance under start-of-life conditions.