MTLSO：一种用于逻辑综合优化的多任务学习方法.pdf

《MTLSO：一种用于逻辑综合优化的多任务学习方法.pdf》由会员分享，可在线阅读，更多相关《MTLSO：一种用于逻辑综合优化的多任务学习方法.pdf（45页珍藏版）》请在三个皮匠报告上搜索。

1、MTLSO:A Multi-Task Learning Approachfor Logic Synthesis OptimizationFaezeh Faez1Raika Karimi1Yingxue Zhang1Xing Li2Lei Chen2Mingxuan Yuan2Mahdi Biparva11Huawei Noahs Ark Lab,Toronto,Canada2Huawei Noahs Ark Lab,Hong Kong,ChinaFaezeh Faez et al.(Huawei Noahs Ark Lab)MTLSO1/45Outline1Motivation2Related

2、 Work3Methodology4Experiments5SummaryFaezeh Faez et al.(Huawei Noahs Ark Lab)MTLSO2/45Motivation(I)Challenges in Logic Synthesis Optimization(LSO):Complexity of Modern ICs:Billions of transistors in modern ICs make manual design infeasible.Traditional heuristic-based methods face limitations in achi

3、evingoptimal results.Machine Learning as a Solution:ML enhances LSO by enabling faster and more accurate predictions 6.Data Scarcity:Limited availability of large,labeled datasets hampers machine learningmodels.Overfitting challenges reduce the generalization and reliability ofpredictions.Inefficien

4、cies in Graph Encoding:Large AIGs with numerous nodes pose challenges for plain GNNs.Treating all nodes with equal importance leads to suboptimalrepresentations.Faezeh Faez et al.(Huawei Noahs Ark Lab)MTLSO3/45Motivation(II)Purpose of MTLSO:Addressing Data Scarcity:Multi-task learning(MTL)enables th

5、e model to leverage sharedsupervision from related tasks.Introducing an auxiliary task(binary multi-label graph classification)enhances model robustness.Improving Graph Representation:Hierarchical graph representation learning captures multi-levelabstractions of AIGs.Combines GNNs with graph downsam

6、pling for better scalability andexpressiveness.Faezeh Faez et al.(Huawei Noahs Ark Lab)MTLSO4/45Related WorkKey Areas in Related Research:Logic Synthesis Optimization(LSO):A growing trend toward employing ML techniques for EDA tasks,moving away from traditional hand-engineered approaches.Techniques