NAND Flash Read Threshold Regression for Lower Bit Error Rates
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
In NAND flash memory, as more bits are stored per memory cell, the threshold voltage window becomes smaller, leading to increased error rates in determining the memory cell's value, and existing error-correction codes struggle to maintain optimal read threshold voltages due to varying operational and storage conditions, making it challenging to achieve low raw bit error rates and improve storage device performance.
Innovation Solution
A computer system employs a regression model with a neural network that determines optimal voltage read thresholds based on operational and storage conditions, using supervised training with a loss function to output voltage read thresholds for each memory page, thereby improving data read accuracy and reducing error rates.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Quantity of substance
If more bits are stored per memory cell, then storage capacity is improved, but the threshold voltage window becomes smaller resulting in increased error rates
Solution Approach 1:
The patent dynamically adjusts the read threshold voltage parameter based on operational conditions (temperature, endurance, retention) and storage conditions (read distribution, die index, block index, wordline index) using a neural network regression model. This allows the system to optimize the threshold voltage window for different multi-level cell configurations, maintaining reliable read operations even as storage capacity increases and voltage windows shrink.
2Speed
If a look-up table is used to store optimal read thresholds, then read speed is improved, but the table size becomes huge and implementation becomes challenging
Solution Approach 1:
Instead of storing a complete look-up table with all possible combinations of operational and storage conditions, the patent uses a neural network regression model that has been trained offline to copy or approximate the optimal read threshold values. The trained model can predict thresholds for any condition combination without requiring explicit storage of all possibilities, dramatically reducing memory requirements while maintaining fast prediction speed.
Solution Approach 2:
The patent transforms the problem from a high-dimensional look-up table approach to a continuous function approximation using a neural network. By training the network on a subset of data and then using it to generalize across the entire input space, the system effectively moves from discrete table storage to continuous mathematical modeling, reducing complexity while preserving accuracy.
3Measurement precision
If offline memory characterization is performed to determine optimal read thresholds, then accuracy is improved, but not all possible combinations and realizations can be determined
Solution Approach 1:
The patent performs preliminary offline training of the neural network regression model using a comprehensive set of operational and storage conditions. During this training phase, the system characterizes memory behavior across various conditions and learns the relationships between input parameters and optimal read thresholds. Once trained, the model can generalize to predict thresholds for conditions not explicitly present in the training data, providing both accuracy and broad adaptability.
Solution Approach 2:
The system implements a feedback mechanism where the neural network continuously receives actual operational and storage conditions as inputs and adjusts its predictions accordingly. The model is trained using measured data that includes feedback about actual memory behavior under different conditions, allowing it to learn from real-world performance and improve its threshold predictions across the full range of possible conditions.
Data Source
AI summary
Techniques related to improving a performance related to at least data reads from a memory are described. In an example, a computer system hosts a regression model that includes a neural network. The neural network is trained based on training data that is measured under different combinations of operational conditions and storage conditions. In operation, actual operational and storage conditions associated with the memory are input to the regression model. The neural network outputs a voltage read threshold based on these actual conditions. The computer system uses the voltage read threshold to read data stored in the memory.