Abstract
NAND flash memory is widely used in various computing systems. However, flash blocks can sustain only a limited number of program/erase (P/E) cycles, which are referred to as the endurance. On one hand, in order to ensure data integrity, flash manufacturers often define the maximum P/E cycles of the worst block as the endurance of flash blocks. On the other hand, blocks exhibit large endurance variations, which introduce two serious problems: over-provisioned error correcting code (ECC) and underutilized block lifespan. In this paper, to exploit the storage capacity of each flash block, we propose an RBER-aware lifetime prediction scheme based on machine learning technologies. We use incremental learning to update the model for adapting the changes at different lifetime stages. Experimental results show that the support vector machine (SVM) models based on our proposed scheme can achieve as high as 95% accuracy. Furthermore, applying this scheme to predict actual endurance at different retention times can improve the maximum P/E cycle of flash blocks from 37.5% to 86.3% on average.
Introduction: The Endurance Bottleneck
NAND flash memory is a consumable resource; every time you program or erase (P/E) data, the physical structure of the memory cell degrades. Traditionally, manufacturers define a "nominal endurance"—a maximum number of P/E cycles—after which a block is considered unreliable. However, this limit is extremely conservative. Because of manufacturing variations, some blocks are much more resilient than others. To ensure data integrity, manufacturers set limits based on the worst-performing block, leading to significant underutilization of the storage medium.
The Problem: Conservative Limits and Underutilization
Setting endurance limits based on the worst-case scenario leads to two major inefficiencies. First, most blocks are retired while they still have significant healthy life remaining. Second, Error Correcting Codes (ECC) are over-provisioned to handle these worst-case scenarios, which increases decoding latency for every read operation, even on healthy blocks.
As shown in the distribution of Raw Bit Error Rates (RBER), there is a wide variance in block health. The authors aim to solve this by predicting the RBER—the actual health metric—of each individual block in real-time using machine learning.
Methodology: Predicting Health in Real-Time
The researchers proposed an RBER-aware lifetime prediction scheme using four machine learning models: Support Vector Machines (SVM), Random Forests (RF), Multi-layer Perceptrons (MLP), and Long Short-Term Memory (LSTM) networks. The core of their approach involves:
- Feature Selection: Using P/E cycles and retention time (how long data sits in the cell) as primary inputs. Retention time is critical because voltage distributions shift over time, as illustrated below.
- Data Smoothing: Utilizing Exponentially Weighted Moving Average (EWMA) to filter out environmental noise in RBER data.
- Incremental Learning: To combat "model aging," the system monitors prediction accuracy ($R^2$). If it drops below 0.9, the model is automatically retrained using the most recent data from the flash controller.
Key Results and Insights
The study utilized a custom test platform with real 512G 3D-TLC NAND flash chips. Key findings include:
- SVM Efficiency: While LSTM models were stable, they were computationally expensive. SVM provided a 95% accuracy rate with significantly lower time overhead, making it the most practical choice for integration into SSD controllers.
- Lifespan Extension: By using predicted endurance instead of nominal limits, the usable P/E cycles improved by an average of 37.5% to 86.3% across various retention scenarios.
- Dynamic Updating: The incremental learning strategy successfully kept the $R^2$ value above 0.9 throughout the entire lifespan, whereas static models saw accuracy plummet as blocks aged physically.
Conclusion
This research demonstrates that moving away from "one-size-fits-all" endurance limits can significantly extend the life of 3D-NAND flash. Future work may involve implementing "Adaptive ECC," where error correction strength is adjusted based on the predicted health of specific blocks, potentially reducing overall system latency. While hardware overhead remains a consideration, the potential for nearly doubling the usable lifespan of flash storage presents a compelling case for ML-integrated memory controllers.
Summary updated