AI Processor Sub-Network Segmentation for Offline Model Compatibility
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current processor-based information processing systems are incompatible with various off-line network models due to the lack of differentiation between network layers, limiting their ability to operate with multiple types of off-line networks.
Innovation Solution
A method and device for constructing off-line network models by identifying and defining operating parameters for each sub-network, allowing for classification and assignment of sub-networks to appropriate processors, enabling compatibility with multiple types of off-line networks and expanding the range of AI processing devices that can run the network models.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If a single processor is used to operate off-line network models without differentiating network layers, then the device structure is simple, but the processor becomes incompatible with various types of off-line network models
Solution Approach 1:
The patent divides the network model into multiple sub-networks with different types (first type, second type, third type), each processed by dedicated processing units. This segmentation allows the processor to handle various network model types without requiring a completely different processor architecture for each type, thus improving compatibility while maintaining manageable complexity.
Solution Approach 2:
The patent creates a universal processor structure that includes multiple types of processing units (first processing unit, second processing unit, third processing unit) that can collectively handle different network model types. Each processing unit is designed to be compatible with specific network layer types, and the combination provides universal compatibility across various off-line network models.
2Adaptability or versatility
If each layer of network is not differentiated in the constructed off-line network model, then the model construction process is simple, but a single processor cannot be compatible with various off-line network models
Solution Approach 1:
The network model is segmented into distinct sub-networks with clearly defined types. Each sub-network type corresponds to specific processing unit types, creating a structured differentiation that enables compatible processing while maintaining model organization and manageability.
Solution Approach 2:
Different parts of the network model (different sub-network types) are assigned different processing unit types based on their specific requirements. This local differentiation ensures that each part of the network model is processed by the most appropriate processing unit, improving overall compatibility and processing efficiency.
3Productivity
If all sub-networks are processed by the same processing unit type, then the processing system is simple to manage, but the processing efficiency and accuracy for different network types are suboptimal
Solution Approach 1:
The patent assigns different processing unit types to different sub-network types based on their specific processing requirements. First processing units handle first type sub-networks, second processing units handle second type sub-networks, and third processing units handle third type sub-networks. This localized optimization improves processing efficiency and accuracy for each network type while maintaining a manageable configuration through clear type correspondence.
Data Source
AI summary
The present disclosure provides a network off-line model processing method, an artificial intelligence processing device and related products, where the related products include a combined processing device. The combined processing device includes the artificial intelligence processing device, a general-purpose interconnection interface, and other processing devices, where the artificial intelligence processing device interacts with the other processing devices to jointly complete computation designated by users. The embodiments of the present disclosure can accelerate the operation of the network off-line model.