Neural Network Operator Fusion for Computation Speed
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing neural network models face inefficiencies in computation speed due to frequent processor switching and data copying when operating on multiple computation nodes, leading to reduced operation speed.
Innovation Solution
A data processing method that acquires neural network operators, fuses them according to a preset rule, combines the fused operators into computation instructions, and uses a computation engine to perform computations, optimizing the processing of data streams by reducing unnecessary processor switching and data copying.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Power
If multiple computation nodes are used to process neural network operations, then computational capability is improved, but processor switching frequency increases and data copying overhead increases, reducing operation speed
Solution Approach 1:
The patent merges multiple computation nodes into a unified processing architecture where operators are fused across what would traditionally be separate processing units. This consolidation eliminates the need for frequent data copying between processors while maintaining the computational capability of multiple nodes, thus resolving the contradiction between computational power and operation speed
2Adaptability or versatility
If neural network operators are processed separately on different processors, then computational diversity is improved, but communication frequency between processors increases, reducing efficiency
Solution Approach 1:
The patent implements a universal processing architecture that can handle diverse neural network operators within a single computation node. This multi-functional design allows the system to maintain computational diversity for different operator types while eliminating inter-processor communication overhead, as all operator types can be processed within the same processing unit
Data Source
AI summary
A data processing method and apparatus for a neural network model, a device, and a storage medium are provided. The method includes: acquiring multiple neural network operators in a neural network model; fusing the multiple neural network operators according to a preset rule to obtain fused neural network operators; combining the fused neural network operators into computation instructions; and performing computation on the computation instructions by using a computation engine.