Contrastive Sample Training for Neural Network Macro Placement
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The high cost and inefficiency of training macro placement tools due to the need for extensive labeling of placement samples, particularly for rare features like 'unusable area,' which requires repeated sample generation and evaluation, making it difficult to integrate with online reinforcement learning.
Innovation Solution
A method using contrastive samples, where positive samples are constructed by sequentially removing macros from a partially-placed chip canvas in different orders, and negative samples by randomly placing macros on an empty canvas, to train a neural network (NN) and graph NN, minimizing labeling costs and improving training efficiency.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If extensive placement samples are generated and labeled to train the neural network for macro placement, then the training accuracy and model performance are improved, but the computational time, resources, and licensing costs increase significantly
Solution Approach 1:
The patent applies preliminary action by pre-processing placement samples to automatically identify and label those containing specific features (e.g., unusable area) before training. This advance preparation eliminates the need for time-consuming manual labeling during the training process, significantly reducing computational time and resources while maintaining training accuracy.
Solution Approach 2:
The system implements self-service through automated feature detection algorithms that enable the training data to label itself. The neural network and associated processing automatically identify samples with problematic features without requiring external expert intervention, thereby reducing both time and licensing costs while preserving measurement precision.
2Manufacturing precision
If manual labeling is performed to identify samples with rare features like 'unusable area,' then the training data quality is improved, but the process becomes extremely time-consuming and resource-intensive
Solution Approach 1:
The patent replaces the mechanical system of manual expert labeling with an automated computational system. Algorithms process placement samples to automatically detect and label features such as unusable areas, substituting human expert review with machine-based detection that maintains data quality while dramatically improving labeling efficiency and reducing resource consumption.
3Adaptability or versatility
If the same feature detection process is repeated for every new type of feature, then comprehensive training coverage is achieved, but the overall training time and resource costs multiply
Solution Approach 1:
The patent implements universality through a single automated feature detection system that can identify multiple different feature types (e.g., unusable area, congestion, wirelength issues) within the same training pipeline. This multi-functional approach eliminates the need to repeat the entire detection process for each feature type, thereby maintaining comprehensive feature coverage while significantly reducing total training time and resource costs.
Data Source
AI summary
A system trains a neural network (NN) for macro placement. The system constructs a set of positive samples of trajectories by sequentially removing the same set of macros in different orders from an at least partially-placed canvas of a chip. The system also constructs a set of negative samples of trajectories by placing not-yet-placed macros at random positions on an at least partially-empty canvas of the chip. The system then trains the NN and a graph NN (GNN) in the NN using the positive samples and the negative samples.