Memory Health Monitoring Using ML for Data Retention Refresh
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current memory sub-systems face challenges in accurately predicting data retention issues due to temporal voltage shift, leading to increased bit error rates and inefficient data refresh operations, which can impact performance and lifespan.
Innovation Solution
A machine-learning-based system health monitoring approach using a TR model that correlates die-level data with voltage offset bins to predict code word error rates, allowing for early identification of data retention issues and optimizing data refresh operations.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional health monitoring methods are used, then system simplicity is maintained, but measurement precision of data retention issues deteriorates
Solution Approach 1:
The patent introduces a machine learning model as an intermediary between raw health scan data and predictive insights. The model processes voltage offset bin assignments and code word error rates to predict future data retention issues, enabling accurate detection without requiring complex hardware modifications. The ML model acts as a software-based mediator that transforms simple scan data into predictive health assessments.
Solution Approach 2:
The system performs preliminary health scans and collects data before actual data retention issues manifest. By continuously monitoring voltage offset bins and code word error rates in advance, the system can predict future failures and trigger preventive data refresh operations, rather than waiting for errors to occur. This preliminary detection approach improves measurement precision while keeping the monitoring system relatively simple.
2Reliability
If frequent data refresh operations are performed, then data retention reliability is improved, but loss of energy increases
Solution Approach 1:
The patent implements a feedback mechanism where the machine learning model continuously analyzes health scan data including voltage offset bin assignments and code word error rates. Based on this feedback, the system dynamically determines when data refresh operations are actually needed, rather than performing frequent refreshes regardless of condition. This feedback-driven approach maintains high data retention reliability while minimizing unnecessary energy-consuming refresh operations.
Solution Approach 2:
Instead of performing full refresh operations on all data blocks frequently, the system applies partial refresh actions only to specific blocks or regions predicted to be at risk. The ML model identifies problematic areas and triggers targeted refresh operations, reducing overall energy consumption while maintaining sufficient reliability. This selective approach avoids excessive refresh operations on healthy data.
3Measurement precision
If comprehensive health scans are performed, then measurement precision of system health is improved, but loss of time increases
Solution Approach 1:
The patent extracts only the most critical health indicators from comprehensive scans, specifically focusing on voltage offset bin assignments and code word error rates. Rather than analyzing all possible health parameters, the system isolates these key metrics that the ML model uses for prediction. This extraction approach maintains measurement precision for detecting data retention issues while significantly reducing the time required for health assessments.
Solution Approach 2:
The system performs preliminary data processing and filtering of health scan results before feeding them to the machine learning model. By pre-processing and organizing health scan data in advance, the system reduces the computational burden during prediction operations. This preliminary organization enables faster, more efficient health assessments without sacrificing the precision of system health evaluation.
4Productivity
If machine learning models are deployed, then productivity of predictive maintenance is improved, but device complexity increases
Solution Approach 1:
The machine learning model is configured to autonomously analyze health scan data and generate predictions without requiring extensive external processing or intervention. The model self-manages the predictive maintenance workflow by automatically interpreting voltage offset bin assignments, calculating code word error rates, and identifying when data refresh operations are needed. This self-service capability improves productivity while minimizing the complexity burden on the host system, as the ML model handles the complex analysis internally.
Data Source
AI summary
A system includes a memory device comprising multiple dies and a processing device coupled to the memory device. The processing device causes a health scan to be performed on a block family of the multiple dies. The processing device updates, for each read sample taken during the health scan, an assignment to a voltage offset bin of a plurality of voltage offset bins. The processing device provides the updated assignments to the plurality of voltage offset bins to a machine learning trigger rate (TR) model. The processing device receives, from the machine learning TR model, a code word error rate (CWER) distribution of the memory device. In response to determining that a TR margin value associated with the CWER distribution does not satisfy a threshold value, the processing device causes a data associated with the block family to be refreshed.