Experimental Results

Comprehensive analysis of machine learning performance in IR-drop aware timing optimization

SVM Performance

Support Vector Machine results for binary classification of timing parameters:

ACCURACY95.45% classification accuracy for Agg_delay_lh
PRECISION96% precision for positive class identification
F1 SCORE97% F1 score demonstrating robust performance
LINEARLinear kernel providing computational efficiency

Random Forest Results

Ensemble learning approach for complex non-linear relationships:

ACCURACY95.80% accuracy for Agg_delay_hl classification
FEATUREFeature importance evaluation for interpretability
ENSEMBLEMultiple decision trees reducing overfitting
NON-LINEARHandles complex non-linear relationships effectively

Key Performance Metrics

Quantified improvements across different threshold values

Classification Accuracy

Both SVM and Random Forest achieved over 95% accuracy across threshold cut-offs from 1% to 5%, demonstrating robust performance in binary classification tasks.

Critical Path Impact

The methodology achieved less than 1% impact on critical path delay while effectively mitigating IR-drop and crosstalk effects through strategic timing optimization.

Timing Optimization

Analysis across threshold cut-offs revealed optimal timing relationships between aggressor and victim gates, enabling efficient timing closure.

Key Findings

Our experimental results demonstrate significant improvements in timing optimization:

High Classification Accuracy

Both algorithms achieved over 95% accuracy in predicting timing parameters

Minimal Performance Impact

Less than 1% impact on critical path delay while mitigating IR-drop effects

Robust Feature Selection

Victim-side propagation delays showed strongest correlations

Optimal Thresholds

Analysis across 1-5% thresholds revealed optimal timing relationships