机器学习在集成电路物理设计中的应用及部署.pdf

1、Use Cases and Deployment of ML in IC Physical Design Amur Ghose,aghoseucsd.eduAndrew B.Kahng,abkucsd.eduSayak Kundu,sakunduucsd.eduYiting Liu,yil375ucsd.eduBodhisatta Pramanik,bopramanikucsd.eduZhiang Wang,zhw033ucsd.eduDooseok Yoon,d3yoonucsd.edu2AI/ML techniques have been applied to many IC physic

2、al design challenges,e.g.:Hyperparameter autotuning for better PPA tool settingsML Predictions of routing hotspots,doomed runs,and PPARouting blockage creation to improve routability and PPABut:practical challenges are seen in ML deployment Why have so many efforts fallen short?This talk:Issues surr

3、ounding data for MLHigh-level principles for deploymentBasic“checklists”for data,models,and use casesContext for MLOps and LLM-based application developmentMotivation3Agenda Data Data Outside vs.Inside IC Design Challenges and Ongoing Efforts(Academia and Industry)ML Deployment Key Performance Indic

4、ators(KPIs)and Checklists Machine Learning Operations(MLOps)and Commoditization Challenges for LLM Deployment LLM EDA:Software Engineering Issues Challenges from EDA Flows4 Data is a core concern in ML for IC designData in PD:Scope,Modalities,Challenges Example ChallengesDiverse IC data typesFormal

5、specsHDLGraphsHierarchiesTabular dataImagesHard for GenAI to interpret Generalization across modalitiesIC design data is costly to produceHuge scale,as well!High-quality public data is scarceUnshareable due to proprietary rightsPDK dataCommercial librariesSoft IP dataEDA vendor dataScarce and propri

6、etary data Data qualityLarger datasets do not guarantee better ML models Common problems:Outdated,stale dataIncomplete coverageRisks such as data poisoning5Many initiatives,contributions to mitigate data scarcity Artificial netlist generators(ANG+),proxy PDKs(ASAP7+)Open-source toolchains(OpenROAD,i