Abstract: Multi-level-cell (MLC) NAND flash memory reliability is significantly impacted by Data Retention Noise (DRN), which shifts voltage distributions over time. Traditional error-correcting codes (ECC) often rely on log-likelihood-ratios (LLRs) derived from read-voltage thresholds. This paper presents a novel optimization strategy for these thresholds, specifically accounting for the constraints of finite block length ECC. We introduce a Cross Iterative Search (CIS) algorithm for scenarios with known channel conditions and a Deep Neural Network (DNN) aided approach for real-world applications where data retention time is unknown. By utilizing voltage histograms as DNN inputs, our method effectively predicts optimal thresholds, leading to enhanced program-and-erase (PE) endurance and reduced read latency.
Introduction: The Reliability Bottleneck in MLC Flash
As NAND flash memory density increases, the physical margins for error decrease. In MLC flash, four distinct voltage states ($s_0, s_1, s_2, s_3$) represent two bits of data. The primary challenge to data integrity is Data Retention Noise (DRN), where electron leakage over time causes these voltage distributions to shift and overlap.
As shown in Figure 1, DRN causes the distributions to move toward lower voltages (leftward shift). To maintain low error rates, the read-voltage thresholds must be dynamically adjusted. However, standard optimization techniques often assume infinite code lengths, which does not reflect the finite block lengths used in practical Low-Density Parity-Check (LDPC) implementations. This gap between theoretical capacity and actual coding rates necessitates a more robust optimization framework.
Methodology: From Analytical Search to Deep Learning
The proposed solution operates in two phases. First, for cases where channel parameters like the number of Program-and-Erase (PE) cycles and Data Retention Time (DRT) are perfectly known, we formulate an optimization problem to minimize the maximum decoding error probability. This is solved using the Cross Iterative Search (CIS) algorithm, which identifies near-optimal thresholds by analyzing the Channel Coding Rate (CCR) under finite block length constraints.
In real-world scenarios, however, DRT is nearly impossible to record accurately for every block. To address this, we employ a Deep Neural Network (DNN) strategy based on a Multi-Layer Perceptron (MLP) architecture. The DNN is trained to map the current state of the flash memory—represented by a voltage histogram—directly to the optimal read thresholds.
A critical aspect of the DNN strategy is the input representation. As illustrated in Figure 6, the system uses non-uniform quantization to generate histograms. Unlike uniform quantization, non-uniform bins can be concentrated in the overlap regions between states, where the most valuable information for threshold placement resides. This allows the MLP (consisting of layers with 512, 256, and 128 neurons) to accurately infer the underlying distribution shifts without explicit knowledge of the data's age.
The system is designed as a "second-tier" defense: if standard decoding fails, the controller activates the DNN to sense the current voltage distribution via a histogram, updates the thresholds, and triggers a second, successful decoding attempt.
Results and Impact
Simulation results demonstrate that the proposed DNN-aided optimization significantly outperforms traditional methods. By accounting for finite block length, the CIS algorithm extends PE endurance to approximately 15,900 cycles, a notable improvement over entropy-based (15,600) or mutual information-based (15,100) approaches.
Furthermore, the DNN-aided scheme exhibits superior tolerance to long retention times. At 6,000 PE cycles, it successfully decodes data with up to 30,000 hours of retention, whereas standard Belief Propagation (BP) decoding fails much earlier. Because the DNN provides a "one-shot" threshold update rather than relying on multiple iterative decoding attempts, it also substantially reduces read latency compared to existing retention-aware iterative decoders.
Conclusion
By bridging the gap between theoretical capacity and practical finite-length ECC, and by leveraging the predictive power of deep learning, this work provides a robust framework for extending the lifetime of MLC flash memory. The use of voltage histograms as a proxy for complex noise states offers a practical path toward more intelligent and resilient flash controllers, balancing hardware complexity with significant gains in longevity and performance.
Summary updated