Weight-sharing neural networks for non-volatile memory read threshold prediction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Non-volatile memory devices face challenges in maintaining data integrity due to voltage level drops caused by noise and interference, leading to read errors and reduced longevity, which existing solutions address with additional read operations that increase latency.
Innovation Solution
Implementing a weight-sharing deep neural network architecture that processes history reads to estimate optimal read thresholds, reducing latency by using previous read attempts to determine accurate read voltage levels.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If additional read operations are performed to account for voltage drop, then data retrieval accuracy is improved, but latency increases
Solution Approach 1:
The system performs preliminary actions by conducting history reads and using neural networks to predict optimal read thresholds before the actual data retrieval operation. The neural network processes historical read data to estimate voltage drops and determines compensated read thresholds in advance, so when the actual read operation occurs, the accurate thresholds are already prepared, eliminating the need for additional corrective read operations and thus reducing latency while maintaining high accuracy
2Duration of action of stationary object
If read threshold compensation is performed to account for degradation, then longevity is improved, but device complexity increases
Solution Approach 1:
The patent replaces traditional mechanical/mathematical voltage drop compensation mechanisms with a neural network-based predictive system. Instead of using complex algorithms to calculate and adjust read thresholds based on program/erase cycle counts and voltage drop models, the system uses trained neural networks that automatically process historical read data and output compensated thresholds. This substitution of neural network intelligence for traditional computational mechanics achieves accurate threshold compensation while simplifying the overall system architecture and reducing the burden on the LDPC decoder
Data Source
AI summary
Devices, systems, and methods for improving the performance of a non-volatile memory are described. An example method includes extracting one or more parameters from each of multiple history reads of a page of the memory device, and processing, using each of multiple first neural networks, a corresponding history read of the multiple history reads to generate a corresponding intermediate vector of multiple intermediate vectors. Each of the multiple first neural networks are trained using measurements from one or more pages of another memory device, and each first neural network is activated upon determining that the one or more parameters for the corresponding history read have been extracted. The method further includes processing, using a second neural network, the multiple intermediate vectors to generate an updated read threshold set, and applying the updated read threshold set to the memory device to retrieve information from the memory device.