Memory Wear Leveling for Neural Network Training
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing memory devices for neural networks face challenges with high costs, scaling difficulties, and leakage power issues, particularly with DRAM, while phase-change memory (PCM) offers alternatives but requires effective wear leveling to prolong service life without excessive data movement and computational complexity.
Innovation Solution
A marching-based wear leveling method that separates unrewritable and rewritable data into a buffer zone, ensuring even distribution of data updates and minimizing harm to memory cells by skipping the buffer zone for rewritable data during the training phase.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Duration of action of stationary object
If wear leveling is implemented to distribute erasures evenly across memory cells, then service life of memory cells is extended, but memory space is consumed and operating complexity increases
Solution Approach 1:
The memory array is segmented into a buffer zone and a non-buffer zone. The buffer zone specifically stores unrewritable data during training phase, while the non-buffer zone stores rewritable data. This segmentation allows the wear leveling mechanism to operate differently on different data types, simplifying the overall control logic while effectively distributing wear across memory cells.
Solution Approach 2:
The system performs preliminary classification of data into unrewritable and rewritable categories before storage. During the training phase, unrewritable data is pre-positioned in the buffer zone, and this classification is maintained throughout training. This preliminary action eliminates the need for complex real-time analysis of data rewrite patterns, reducing operational complexity while maintaining effective wear distribution.
2Reliability
If data is moved and redistributed for wear leveling, then even distribution of erasures is achieved, but latency increases due to excessive data searching and moving
Solution Approach 1:
Different quality policies are applied to different zones: the buffer zone receives only unrewritable data with guaranteed wear protection, while the non-buffer zone handles rewritable data with standard wear leveling. This local differentiation eliminates the need for system-wide data movement and searching, reducing latency while maintaining reliable wear distribution in the critical buffer zone.
Solution Approach 2:
Data is classified and positioned in appropriate zones before the training phase begins. This preliminary positioning ensures that unrewritable data is already in the buffer zone when training starts, eliminating the need for time-consuming data searching and moving operations during training execution.
3Reliability
If buffer zone is used to store unrewritable data, then wear to memory cells is reduced, but memory space is consumed by the buffer zone
Solution Approach 1:
The buffer zone serves multiple functions: it protects unrewritable data from excessive wear during training phase, maintains data integrity for neural network training operations, and can be configured to accommodate different training workloads. This multi-functionality justifies the memory space allocation by providing comprehensive wear protection for critical training data.
Solution Approach 2:
The size and configuration of the buffer zone can be adjusted as a parameter to balance between wear protection and memory space availability. By changing the buffer zone size parameter, the system can adapt to different training requirements and memory capacity constraints, optimizing the trade-off between reliability and resource consumption.
Data Source
AI summary
A memory device includes: a memory array used for implementing neural networks (NN); and a controller coupled to the memory array. The controller is configured for: in updating and writing unrewritable data into the memory array in a training phase, marching the unrewritable data into a buffer zone of the memory array; and in updating and writing rewritable data into the memory array in the training phase, marching the rewritable data by skipping the buffer zone.