Machine Learning Threshold Voltage Identification for Memory Read Accuracy
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing memory devices with newer architectures like 3D NAND and TLC/QLC flash memories face issues with inaccurate threshold voltage during read operations, leading to data retention problems, read disturb errors, and increased ECC decoding difficulties, which current management mechanisms fail to adequately address without introducing side effects.
Innovation Solution
A method employing machine learning to generate threshold voltage identification models based on a predetermined database, allowing for accurate bit information retrieval from memory cells by adjusting reference voltages during read operations, thereby improving threshold voltage control without significantly increasing costs.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If machine learning-based threshold voltage identification is implemented, then read operation accuracy is improved, but device complexity increases
Solution Approach 1:
The patent applies preliminary action by pre-training machine learning models during a training phase before actual read operations occur. The controller is pre-configured with threshold voltage identification models that are generated in advance, so that during normal operation, the system only needs to execute the pre-computed models rather than performing complex real-time learning, thus reducing operational complexity while maintaining high accuracy.
Solution Approach 2:
The patent uses copying by creating a simplified representation of the complex threshold voltage distribution through machine learning models. Instead of directly managing the full complexity of threshold voltage variations in each memory cell, the system copies the essential characteristics into compact ML models that can be efficiently executed, reducing the computational burden during read operations.
2Device complexity
If traditional management mechanisms are used, then device complexity is kept low, but read operation reliability deteriorates
Solution Approach 1:
The patent replaces traditional mechanical/voltage-based threshold management with machine learning-based identification. Instead of using fixed voltage thresholds or simple lookup tables, the system substitutes these with ML models that can adapt to complex threshold voltage distributions, thereby improving reliability without requiring overly complex hardware management mechanisms.
Solution Approach 2:
The patent implements feedback by using the trained machine learning models to continuously improve threshold voltage identification based on actual read operation results. The system can learn from read errors and adjust its models accordingly, creating a feedback loop that improves reliability over time while keeping the base system relatively simple.
3Quantity of substance
If higher density memory cells (MLC, TLC, QLC) are used, then storage capacity is improved, but threshold voltage accuracy deteriorates
Solution Approach 1:
The patent applies parameter changes by using machine learning models to dynamically adjust threshold voltage parameters based on the specific characteristics of each memory cell. Instead of using fixed voltage levels that work for SLC but fail for higher density cells, the system changes the threshold voltage parameters adaptively using ML-based identification, enabling accurate reading of TLC and QLC cells while maintaining high storage capacity.
Solution Approach 2:
The patent implements local quality by tailoring threshold voltage identification to the specific characteristics of each memory cell or cell group. Rather than using a one-size-fits-all approach, the system adjusts the identification parameters locally based on the specific cell's threshold voltage distribution, which is crucial for high-density cells where each cell has unique characteristics that affect read accuracy.
Data Source
AI summary
A method for performing memory access management with aid of machine learning in a memory device, the associated memory device and the controller thereof, and the associated electronic device are provided. The method may include: in the memory device, during a training phase, performing machine learning according to a predetermined database regarding threshold voltage distribution, to generate at least one threshold voltage identification model, wherein the at least one threshold voltage identification model is utilized for determining bit information read from a memory cell of the NV memory; and in the memory device, during an identification phase, obtaining representative information of one or more reference voltages when reading the NV memory, for performing machine identification according to the at least one threshold voltage identification model to generate read data, wherein the read data includes the bit information.