Cross-Network Semiconductor Design for Routability Prediction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional semiconductor design methods face challenges in efficiently predicting routability due to the limitations of using either graph-based or image-based neural networks, which fail to effectively capture spatial relationships and result in unnecessary routing overload or information loss.
Innovation Solution
A semiconductor design method utilizing a fusion of graph-based and image-based neural networks, where features are converted and fused across layers to enhance the prediction of routability, minimizing delays and overheads by encoding chip design information in both graph and image formats.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If only graph-based neural network is used for routability prediction, then spatial relationship capture is improved, but information loss occurs in spatial patterns
Solution Approach 1:
The patent combines graph-based neural network and image-based neural network into a hybrid architecture. The graph-based component captures spatial relationships between circuit elements, while the image-based component preserves spatial patterns and visual features. The fusion of features from both networks resolves the contradiction by ensuring neither spatial relationship capture nor spatial pattern information is lost.
2Loss of information
If only image-based neural network is used for routability prediction, then spatial patterns are preserved, but spatial relationship understanding is insufficient
Solution Approach 1:
The hybrid neural network architecture merges the strengths of both image-based and graph-based approaches. The image-based neural network preserves spatial patterns and visual features, while the graph-based neural network provides accurate spatial relationship understanding through node and edge representations. Feature fusion integrates both types of information to achieve comprehensive routability prediction.
3Measurement precision
If separate graph-based and image-based models are used, then comprehensive feature extraction is achieved, but device complexity increases
Solution Approach 1:
The patent introduces a feature fusion module as an intermediary between the graph-based neural network and image-based neural network. This fusion module integrates features from both networks in a structured manner, managing the complexity of the dual-network architecture. The fusion module consolidates the comprehensive feature extraction capabilities while providing a unified interface for routability prediction, thereby controlling system complexity.
4Measurement precision
If comprehensive design data processing is performed, then prediction accuracy is improved, but processing time and overhead increase
Solution Approach 1:
The patent performs preliminary feature extraction and processing in the graph-based and image-based neural networks before fusion. By pre-processing design data into meaningful features in each specialized network, the system reduces the computational burden on the fusion module and final prediction stage. This preliminary action enables comprehensive data processing while minimizing processing time and overhead through efficient feature representation.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
A semiconductor design method and device are provided. The semiconductor design method may include inputting a first type of design data into a first neural network model; inputting a second type of design data into a second neural network model of a different type from the first neural network model; generating a fusion feature by fusing a calculation result of the second neural network model with a feature generated by calculation up through (and obtained from) a first layer of the first neural network model; inputting the fusion feature into a second layer of the first neural network that is after the first layer of the first neural network model; and performing a task related to routability of a circuit after calculation of the first neural network model based on the fusion feature is completed.