Memory System Aging Compensation via Block Segmentation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Computing-in-memory (CiM) operations performed by non-volatile memory devices suffer from accuracy degradation due to aging, requiring calibration to maintain computation accuracy.
Innovation Solution
A memory system with a control circuit that divides the memory array into multiple blocks, enabling or disabling blocks based on aging conditions to compensate for drifted weights by adding or cutting memory blocks, thereby maintaining computation accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Speed
If computing-in-memory operations are performed by non-volatile memory devices, then computation speed is accelerated, but accuracy degrades due to aging
Solution Approach 1:
The patent applies preliminary action by training backup neurons in advance to compensate for expected weight drifts caused by aging. Before the memory device is fully deployed, multiple backup neurons are trained with different compensation parameters. When aging occurs, the pre-trained backup neurons can be activated to compensate for the drifted weights without requiring real-time retraining or calibration, thus maintaining computation accuracy while preserving the speed benefits of CiM operations.
2Reliability
If calibration is performed to maintain computation accuracy, then accuracy is maintained, but additional time and complexity are required
Solution Approach 1:
The patent performs calibration in advance during the training phase, where multiple backup neurons are trained with different compensation parameters before deployment. This preliminary calibration eliminates the need for time-consuming calibration operations during actual computation, as the compensation mechanisms are already prepared and can be activated instantly when aging effects are detected.
Solution Approach 2:
The patent creates copies of neurons (backup neurons) that replicate the functionality of original neurons but with compensated weights. These backup neurons serve as pre-prepared alternatives that can replace drifted neurons without requiring complex real-time calibration procedures, thus maintaining accuracy while minimizing time loss.
3Reliability
If backup neurons are added to compensate for aging, then accuracy is maintained, but device complexity increases
Solution Approach 1:
The patent segments the neural network into original neurons and separate backup neurons, allowing independent management and activation. This segmentation enables the system to use only the necessary components (original neurons when fresh, backup neurons when aging occurs) rather than maintaining all components active simultaneously, thus managing complexity more effectively while maintaining accuracy through selective activation of compensation mechanisms.
Data Source
AI summary
A training method, an operating method and a memory system are provided. The operating method comprises using a first memory block of the memory system for computation; obtaining an aging condition of the memory system; determining whether the aging condition meets a predetermined aging condition; and when it is determined that the aging condition meets the predetermined aging condition, enabling the second memory block and using the first memory block and the second memory block for computation.