Compute-in-Memory Neural Networks for Predictive NVM Block Health
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing non-volatile memory devices suffer from inefficient error management techniques that are either time-consuming or inaccurate in identifying and retiring bad blocks, often leading to data loss due to overkill or underkill.
Innovation Solution
Implementing a pre-trained neural network on the memory device itself for predictive block health assessment, utilizing compute-in-memory technology to analyze operational errors and proactively identify and retire potential bad blocks.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If neural networks are used for predictive block health assessment, then measurement precision and reliability are improved, but device complexity increases
Solution Approach 1:
The patent merges the neural network computational units directly into the memory device structure, combining storage and processing functions. The neural network is implemented using memory cells arranged in arrays where weights are stored in memory locations, allowing the memory device to perform both data storage and predictive health assessment without requiring a separate external processor.
Solution Approach 2:
The memory device is designed to serve multiple functions: it stores user data, stores the neural network weights, performs computational operations for health assessment, and manages block-level operations. This multi-functionality reduces the need for separate dedicated components and integrates the neural network capability within the existing memory architecture.
2Productivity
If compute-in-memory is implemented, then productivity is improved by reducing data transfer, but device complexity increases
Solution Approach 1:
The computational operations for neural network inference are performed directly within the memory array itself rather than requiring data to be transferred to an external processor. The memory cells are configured to perform multiply-accumulate operations during readout, enabling compute-in-memory and improving computational efficiency by eliminating data transfer bottlenecks.
Solution Approach 2:
The patent introduces intermediate computational units within the memory architecture that facilitate the neural network operations. These intermediate units, implemented using memory cell arrangements, serve as mediators between data storage and the final health assessment output, enabling complex computations to be performed within the memory system without requiring external intervention.
3Reliability
If existing error management techniques are used, then device complexity is reduced, but reliability deteriorates due to inaccurate bad block identification
Solution Approach 1:
The neural network continuously monitors and assesses block health status in advance, identifying potential bad blocks before they cause data loss. By performing predictive assessment rather than reactive error correction, the system can proactively manage deteriorating blocks and prevent failures before they occur, improving reliability through early detection and prevention.
Solution Approach 2:
The system implements continuous feedback loops where the neural network monitors operational parameters, assesses block health, and provides feedback for dynamic block management decisions. This feedback mechanism enables the system to adapt to changing block conditions and make real-time decisions about data migration and block retirement, improving reliability through continuous monitoring and adaptive response.
Data Source
AI summary
Predictive block health assessment of a non-volatile memory can be performed by storing a pre-trained neural network for this purpose on the memory device itself. As operational errors occur during operation of the memory device, this defect data is saved by the non-volatile memory and used as input data for the neural network to identify and retire potential bad blocks before data loss occurs.