Machine Learning RBER Estimation for Memory Subsystems
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Traditional memory subsystems incur high latency and power consumption due to the need for error correcting code (ECC) decoding operations to determine the raw bit error rate (RBER), which can block other operations and contribute to increased power usage even when the RBER does not merit media management actions.
Innovation Solution
Implementing a machine learning-based RBER estimator that predicts RBER using features such as population data and location of memory cells, allowing for the deferral of costly ECC decode operations and media management actions until a predicted RBER exceeds a threshold.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If ECC decoding operations are performed to determine the raw bit error rate, then measurement precision of RBER is improved, but use of energy and latency increase
Solution Approach 1:
The patent introduces an intermediary RBER estimation mechanism that uses population data from memory cells as a proxy indicator. Instead of directly performing full ECC decoding to measure RBER, the system uses the distribution of read voltages across memory cell populations to estimate RBER. This intermediary approach provides sufficient measurement precision for media management decisions while consuming significantly less power and introducing minimal latency.
2Measurement precision
If ECC decoding operations are performed to determine the raw bit error rate, then measurement precision of RBER is improved, but productivity decreases due to blocked operations
Solution Approach 1:
The patent employs an intermediary estimation approach using population data that does not block productive operations. By analyzing the statistical distribution of read voltages across memory cell populations, the system可以获得 RBER estimates without halting the data pipeline for full ECC decoding. This maintains measurement precision while allowing concurrent memory operations to proceed, thereby preserving throughput.
Solution Approach 2:
The patent applies partial action by performing only the necessary subset of operations to obtain RBER estimates. Instead of completing full ECC decoding which is excessive for estimation purposes, the system performs partial analysis of population data that provides sufficient information for media management decisions. This partial approach reduces computational overhead and eliminates blocking, maintaining productivity while achieving the required measurement precision.
3Reliability
If ECC decoding operations are performed to determine the raw bit error rate, then reliability of media management decisions is improved, but use of energy increases
Solution Approach 1:
The patent uses population data as an intermediary indicator that correlates with actual RBER. By analyzing the distribution characteristics of read voltages across memory cell populations, the system derives RBER estimates that are reliable enough for media management decisions such as read voltage calibration and data retention monitoring. This intermediary approach achieves the required decision reliability while consuming a fraction of the power needed for full ECC decoding operations.
Data Source
AI summary
Techniques for estimating raw bit error rate of data stored in a group of memory cells are described. Encoded data is read from a group of memory cells. A first population value is obtained based on a first number of memory cells in the group of memory cells having a read voltage within a first range of read voltages, each read voltage representing one or more bits of the encoded data. An estimated raw bit error rate of the data is determined to satisfy a first threshold. The determination is made using a first trained machine learning model and based in part on the first population value. A first media management operation is initiated in response to the determination that the estimated raw bit error rate satisfies the first threshold.