Gradient Descent Read Threshold Generation for Solid State Storage
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing solid state storage systems, such as NAND Flash, require selection of read thresholds without side information, which is inefficient as the number of bits and read thresholds increase, and often take longer to select next thresholds, consuming significant overhead information and resources.
Innovation Solution
A gradient descent read threshold generation technique that determines next read thresholds using gradients without side information, allowing for efficient selection across various solid state storage systems, including SLC, MLC, and TLC systems, by performing multiple reads and calculating gradients to adjust thresholds dynamically.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If side information is stored to assist in selecting read threshold, then read threshold selection accuracy is improved, but overhead information consumption increases significantly
Solution Approach 1:
The patent extracts and eliminates the need for side information storage by using gradient descent algorithms that operate solely on read data. The system performs multiple reads at different thresholds and uses the gradient information from read results to dynamically adjust thresholds without requiring any additional side information to be stored in the memory device.
Solution Approach 2:
The read threshold selection process becomes self-service by using the read results themselves to generate gradient information that drives threshold adjustment. The system serves its own need for threshold optimization without external assistance from pre-stored side information, making the process autonomous and eliminating overhead.
2Measurement precision
If traditional threshold selection techniques are used, then read threshold can be determined, but processing time increases significantly as number of bits and read thresholds increase
Solution Approach 1:
The patent implements dynamic threshold adjustment through gradient descent, where read thresholds are continuously optimized based on real-time gradient calculations from read results. This dynamic approach replaces static, pre-determined thresholds with adaptive thresholds that automatically adjust to changing data conditions, significantly reducing processing time as the system scales.
Solution Approach 2:
The system changes the parameter approach by using gradient-based optimization instead of traditional fixed or lookup-table methods. The gradient descent algorithm iteratively adjusts threshold parameters based on calculated gradients from read data, enabling efficient scaling to higher bit counts and multiple read thresholds without linear increases in processing time.
3Productivity
If gradient descent method is used to generate read thresholds, then processing efficiency and scalability are improved, but complexity of threshold generation process increases
Solution Approach 1:
The gradient descent method implements feedback mechanisms where read results are used to calculate gradients that feed back into threshold adjustment. This closed-loop feedback system automatically optimizes thresholds based on actual read performance, simplifying the overall process despite the underlying mathematical complexity by using iterative refinement rather than complex one-time calculations.
Data Source
AI summary
A first bit position of a cell in solid state storage is read where a sorting bit is obtained using the read of the first bit position. A second bit position of the cell is read for a first time, including by setting a first read threshold associated with the second bit position to a first value and setting a second read threshold associated with the second bit position to a second value. The second bit position of the cell is read for a second time, including by setting the first read threshold to a third value and setting the second read threshold to a fourth value. A new value for the first read threshold and for the second read threshold is generated using the sorting bit, the first read, and the second read.