Memory Index Mapping for DNN Fault Tolerance
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Deep neural network operations require large memory space to store weights, and stuck-at-faults in memory can lead to incorrect operations, necessitating a solution to reduce faults and improve accuracy.
Innovation Solution
A memory system with a processing unit and a weight unit that includes an index memory and a mapping table, where weight indexes are mapped to representative weight data to reduce memory usage and detect and minimize stuck-at-faults by selecting coded data with the least faults.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Volume of stationary object
If memory space is reduced by grouping weight values into representative data, then memory capacity requirements are decreased, but the risk of stuck-at-faults affecting operation accuracy increases
Solution Approach 1:
The patent segments the weight storage into two parts: an index memory storing compact weight indexes and a mapping table storing the mapping relationships. This segmentation allows the system to reduce memory space while maintaining reliability by separating the compact storage function from the mapping function, enabling fault detection and correction mechanisms to operate on the mapping table without affecting the compactness of weight storage.
Solution Approach 2:
The patent introduces a mapping table as an intermediary between the index memory and the actual weight values. This mapping table acts as a buffer that can detect and correct stuck-at-faults, preventing them from directly affecting the neural network operation accuracy while maintaining the space efficiency of index-based storage.
2Reliability
If mapping table is created by detecting faults and selecting coded data with least stuck-at-faults, then operation accuracy is improved, but the complexity of memory initialization increases
Solution Approach 1:
The patent performs fault detection and mapping table creation as preliminary actions during system initialization or manufacturing. By detecting stuck-at-faults in advance and creating an optimized mapping table before actual operation, the system eliminates the need for complex real-time fault handling during neural network computation, thus improving operational accuracy without significantly impacting runtime performance.
Solution Approach 2:
The system performs self-diagnosis by automatically detecting stuck-at-faults in the index memory and autonomously generating an optimized mapping table that avoids faulty locations. This self-service approach eliminates the need for external manual configuration or complex external fault handling mechanisms, balancing improved reliability with acceptable initialization complexity.
Data Source
AI summary
A memory is suitable for performing a deep neural network operation. The memory includes: a processing unit and a weight unit. The processing unit includes a data input terminal and a data output terminal. The weight unit is configured to be coupled to the data input terminal of the processing unit. The weight unit includes an index memory and a mapping table. The index memory is configured to store multiple weight indexes. The mapping table is configured to respectively map the multiple weight indexes to multiple representative weight data.