Multi-Accelerator Feature Map Processing via Shift Register Arrays
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing neural network accelerator technologies face high memory storage demands and NOC bandwidth waste due to overlapping feature data handling methods, leading to inefficient processing of convolution operations.
Innovation Solution
Implementing shift register arrays to transfer overlapping feature data between accelerators, allowing each to store non-overlapping data, thereby reducing storage requirements and NOC bandwidth usage.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If the feature data of the overlap area is stored in the memory of each accelerator, then each accelerator can independently complete its operation task, but the storage space demand for the memory in the accelerator increases
Solution Approach 1:
The patent introduces an overlap buffer as an intermediary component between accelerators. The overlap buffer stores the overlap feature data and is shared among multiple accelerators through a memory interconnection network, eliminating the need for each accelerator to maintain its own copy of overlap data while enabling independent operation.
2Quantity of substance
If the feature data of the overlap area is transmitted between memories of various accelerators through NOC communication, then the storage space demand is reduced, but the NOC bandwidth is wasted due to data transmission
Solution Approach 1:
The patent merges the overlap feature data storage function into a shared overlap buffer that is collectively accessed by multiple accelerators. This consolidation eliminates redundant data transmission over the NOC, as the overlap data resides in a shared location that all accelerators can access without requiring repeated transmissions.
3Power
If a plurality of accelerators are used to process the same feature map, then the computational power is increased, but the data transmission between accelerators increases NOC bandwidth consumption
Solution Approach 1:
The patent segments the feature map processing task among multiple accelerators while introducing an overlap buffer to handle boundary conditions. Each accelerator processes a specific region of the feature map, and the overlap buffer efficiently manages the boundary data that needs to be shared, reducing the overall data transmission requirements compared to a fully distributed approach.
Data Source
Figure 1~2
Figure 3
Figure 4
AI summary
Disclosed are a method and apparatus for processing neural network feature map using a plurality of accelerators. The method includes: reading first feature data about the neural network feature map from first shift register array in first accelerator among a plurality of neural network accelerators, and first weight data corresponding to the first feature data from first buffer; performing preset operation on the first feature data and first weight data using the first accelerator, to obtain a first operation result; shifting, according to preset shift rule, first overlapping feature data in the first feature data and required by a second accelerator to a second shift register array of the second accelerator; and performing a preset operation on the second feature data from the second shift register array including the first overlapping feature data and the read second weight data using the second accelerator, to obtain a second operation result.