ML-Based DRV Prediction for Advanced Node Design Closure
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Advanced technology nodes face challenges in closing the gap between global route and detail route due to high miscorrelation, leading to performance, power, and area (PPA) closure issues, with conventional algorithms struggling to predict design rule violations (DRVs) efficiently, especially at nodes like 7 nm and below.
Innovation Solution
A machine learning (ML) system is trained to predict design rule violations by utilizing features from both global and detailed routing stages, including buried nets, pin proximity, and power distribution networks, to provide early prediction and reduce turnaround time in the design flow.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If conventional rule-based algorithms are used to predict design rule violations, then the prediction process is simple and fast, but the prediction accuracy is low due to high miscorrelation between global route and detail route
Solution Approach 1:
The patent replaces conventional rule-based algorithms with a machine learning system that uses neural networks to predict design rule violations. This substitution enables the system to learn complex patterns and correlations from training data, significantly improving prediction accuracy while maintaining computational efficiency through the trained model's inference capability.
Solution Approach 2:
The patent implements early prediction of design rule violations at the global route stage using the trained machine learning system. By performing predictions before detailed routing, the system enables designers to identify and resolve potential violations upfront, avoiding costly redesign cycles later in the flow.
2Measurement precision
If detailed route is performed to accurately identify design rule violations, then prediction accuracy is high, but runtime overhead is prohibitively high
Solution Approach 1:
The patent creates a trained machine learning model that copies the predictive capabilities of detailed route analysis into a lightweight inference system. The model is trained on data from detailed routing results but can then predict violations at the faster global route stage, providing accurate predictions without the computational overhead of full detailed routing.
Solution Approach 2:
The system performs preliminary prediction of design rule violations at the global route stage using the trained machine learning model. This early identification allows designers to address potential issues before committing resources to detailed routing, significantly reducing overall runtime while maintaining high accuracy through the model's learned patterns.
3Reliability
If P&R tools are tailored for each customer to achieve optimum PPA, then design optimization is improved, but development time and resource consumption increase significantly
Solution Approach 1:
The patent implements a universal machine learning system that can be trained on customer-specific data and then applied to predict design rule violations across multiple designs and processes. The system maintains a training mode for customization and an inference mode for efficient prediction, allowing customers to optimize PPA for their specific requirements while reusing the same underlying platform and model architecture.
Data Source
AI summary
A machine learning (ML) system is trained to predict the number of design rules violations of a circuit design that includes a multitude of Gcells. To achieve this, a netlist associated with the circuit design is placed by a place and route tool. A first list of features associated with the placed netlist is delivered to the ML system. A global route of the circuit design is performed by a global router. Next, a second list of features is delivered from the global router to the ML system. Thereafter, a detailed route of the circuit design is performed by a detailed router. A label associated with each Gcell in the circuit design is delivered to the ML system from the detailed route. The ML system is trained using the first and second list of features and the labels.