Hard Macro Placement QoR Prediction From Related Circuit Designs
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The challenge in integrated circuit design is the late calculation of quality of results (QoR) metrics for macro cell placements, which leads to significant rework if poor placements are identified only after completing the physical design flow, and existing machine learning approaches struggle with diverse and limited training data.
Innovation Solution
A discerning ensemble of machine learning models is used to predict QoR metrics for candidate macro placements, trained on closely related circuit designs, with predictions combined based on the applicability of the source designs to the target design, utilizing pre- and post-processing to enhance accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If QoR metrics are calculated only later in the physical design flow, then calculation accuracy is improved, but macro placement rework increases significantly
Solution Approach 1:
The patent applies preliminary action by training machine learning models in advance on completed macro placements from previous designs. These pre-trained models can then predict QoR metrics during the macro placement stage itself, providing early feedback without requiring completion of the entire physical design flow. This resolves the contradiction by enabling accurate QoR estimation early in the process, preventing rework while maintaining measurement precision.
2Adaptability or versatility
If machine learning models are trained on diverse circuit designs, then model generalizability is improved, but training data quality and applicability deteriorate
Solution Approach 1:
The patent applies local quality by evaluating the applicability of each trained model to the specific target design and weighting predictions accordingly. Instead of uniformly applying all models regardless of relevance, the system assesses how well each model's training data matches the current design characteristics and gives higher weight to more applicable models. This resolves the contradiction by maintaining generalizability through diverse training while ensuring prediction accuracy through localized applicability assessment.
3Reliability
If multiple machine learning models are used to predict QoR metrics, then prediction reliability is improved, but computational complexity increases
Solution Approach 1:
The patent applies parameter changes by dynamically adjusting the weighting parameters of different models in the ensemble based on their applicability to the target design. The system calculates applicability scores and uses these to determine optimal weights for each model's prediction, rather than using fixed equal weighting. This resolves the contradiction by improving prediction reliability through multiple models while managing complexity through parameter-based adaptive weighting rather than requiring complex integration of all models equally.
Data Source
AI summary
In one aspect, QoR metrics for different candidate macro placements are estimated using machine learning models. A set of candidate macro placements of hard macros within a circuit design is assessed by estimating a quality metric for each candidate macro placement, as follows. Model-specific estimates of the quality metric are predicted by applying different machine learning models to the candidate macro placements. The different machine learning models are trained using sets of completed macro placements for other circuit designs. The model-specific estimates of the quality metric are combined based on an applicability of (a) the set of macro placements used to train that model to (b) the set of candidate macro placements being evaluated.