Neural Network Feature-Map Memory Mapping for SRAM Bank Conflicts
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The storage of feature maps in static random-access memory (SRAM) can lead to bank conflicts, limiting data transmission size per cycle and necessitating additional cycles for neural network operations.
Innovation Solution
A memory mapping method that shifts the position of data within the feature map based on neural network parameters, using techniques like padding and interval shifting, to optimize memory bandwidth utilization.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of manufacture
If feature map is stored in linear address space in SRAM, then storage simplicity is improved, but bank conflicts occur limiting data transmission size
Solution Approach 1:
The feature map data is segmented and distributed across different memory banks using a non-linear address mapping function. The mapping function divides the linear address space into segments that are strategically assigned to different banks, preventing simultaneous access conflicts while maintaining efficient data transmission capability.
Solution Approach 2:
The patent transforms the one-dimensional linear address space into a multi-dimensional memory bank assignment space. By introducing additional mapping dimensions through the non-linear function, data can be accessed across multiple banks simultaneously, resolving the bank conflict limitation while preserving storage simplicity.
2Ease of manufacture
If feature map is stored in linear address space in SRAM, then storage simplicity is improved, but additional cycles are required for processing
Solution Approach 1:
The non-linear address mapping is pre-configured and established before data processing begins. By preliminarily organizing the feature map data according to the mapping function, the system eliminates the need for additional processing cycles during actual neural network operations, as data is already positioned for optimal parallel access across memory banks.
3Productivity
If data position is shifted based on memory bandwidth, then memory bandwidth utilization is improved, but memory access complexity increases
Solution Approach 1:
The patent employs a non-linear mapping function that dynamically adjusts address parameters based on memory bandwidth requirements and kernel size. By changing the address mapping parameters according to operational conditions, the system optimizes bandwidth utilization while the function itself manages the complexity through mathematical transformation rather than complex control logic.
Data Source
AI summary
A memory mapping method includes storing a feature map including a plurality of sets of data used for a neural network operation in a memory, shifting a position of a portion of the data included in the feature map that is stored in the memory based on a parameter of the neural network operation, and outputting data requested by the neural network operation from the feature map in which the position of the portion is shifted based on a memory bandwidth.