Flash Memory Decoding via Inter-Cell Interference Learning
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Densely packed NAND flash memory cells experience significant inter cell interference (ICI) due to voltage leakage from adjacent cells, which corrupts bit values and reduces storage capacity without increasing die size.
Innovation Solution
A decoding technique that leverages inter cell interference by training a decoder, such as a convolutional neural network, to learn weights from adjacent cells, allowing for reliable recovery of true bit values during read operations, and periodically retraining to adapt to changing interference patterns.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Quantity of substance
If memory cells are densely packed to increase storage capacity, then storage capacity increases, but inter cell interference corrupts bit values and reduces reliability
Solution Approach 1:
The patent applies machine learning decoders that are specifically trained to recognize and correct errors caused by inter-cell interference. The system converts the harmful ICI effect into a learnable pattern, where the decoder learns the interference characteristics during training and can subsequently compensate for them during normal operation, thereby maintaining high storage capacity while improving bit value accuracy
Solution Approach 2:
The patent changes the operational parameters by introducing trained machine learning models with learned weights and thresholds. Instead of using fixed decoding thresholds, the system dynamically adjusts decision parameters based on patterns learned during training, allowing it to operate reliably in the high-density regime where traditional fixed-threshold methods fail
2Reliability
If memory cells are spaced out to reduce inter cell interference, then bit value accuracy improves, but storage capacity decreases
Solution Approach 1:
Rather than spacing out cells to avoid interference, the patent embraces the interference as a learnable phenomenon. The machine learning decoder is trained on data that includes the characteristic interference patterns, allowing it to compensate for ICI effects and achieve high accuracy even with densely packed cells, thereby maintaining both high storage capacity and bit value accuracy
3Productivity
If a decoder is trained during manufacturing to learn interference patterns, then read operation performance improves, but device complexity increases
Solution Approach 1:
The patent performs the complex training operation during manufacturing or initialization, before the device enters normal operation. The machine learning decoder is pre-trained on training data that captures the interference characteristics of the specific memory device. This preliminary action transfers the complexity burden to the manufacturing phase, allowing the device to operate efficiently during normal use with the trained model already in place
Solution Approach 2:
The system performs self-characterization and self-training by using its own read operations to generate training data. The decoder learns from the actual interference patterns present in the device, making the training process adaptive to the specific hardware characteristics without requiring external calibration equipment or complex manual procedures
Data Source
AI summary
A method of decoding data from high density memory includes reading voltage levels from a first memory cell and a set of one or more neighboring memory cells of flash memory in response to a read command with an address corresponding to the first memory cell, inputting the voltage levels into a trained model that has been trained on the flash memory to estimate bit values written to a memory cell based on respective voltage values read from the first memory cell and from the neighboring memory cells according to a layout of memory cells of the flash memory, obtaining from the trained model an estimated bit value written to the first memory cell based on the respective voltage levels of the first memory cell and the neighboring memory cells having been input into the trained model, and outputting the estimated hit value in response to the read command.