Towards Collaborative Intelligence: Routability Estimation based on Decentralized Private Data
TimeThursday, July 14th10:52am - 11:15am PDT
Location3007, Level 3
Physical Design and Verification, Lithography and DFM
DescriptionApplying machine learning (ML) is popular in EDA applications from design prediction to optimization.
However, its effectiveness largely hinges on the availability of a large amount of latest design data, which is company-owned and mostly confidential to EDA developers.
Besides, data within a single company might still be inadequate and biased for ML training.
We propose an approach based on federated learning where ML models can be trained with data from multiple clients without violating data confidentiality.
The results are demonstrated through routability prediction.
Furthermore, we customize the routability model considering the adaptation to federated learning.
Experiment results show that collaborative training shows 16% higher accuracy, and our customized model outperforms the best of previous routability estimators by 3.3%.