基于强化学习和半并发优化的混合尺寸布局原型.pdf

1、The EDA LabNational Taiwan UniversityMixed-Size Placement Prototyping Based on Reinforcement Learning with Semi-Concurrent OptimizationCheng-Yu Chiang,Yi-Hsien Chiang,Chao-Chi Lan,Yang Hsu,Che-Ming Chang,Shao-Chi Huang,Sheng-Hua Wang,Yao-Wen Chang,and Hung-Ming ChenASPDAC25,January 2023,2025,Tokyo,J

2、apanThe EDA LabOutline Introduction Problem Formulation Proposed Approach Experimental Results Concluding RemarksGIEE,NTU2The EDA LabOutline Introduction Problem Formulation Proposed Approach Experimental Results Concluding RemarksGIEE,NTU3The EDA LabCircuit Placement Based on Machine LearningPlace

3、objects(macros and standard cells)into a die s.t.no objects overlap with each other&some cost metric(e.g.,wirelength&power)is optimizedCritical and time-consuming stage that greatly affects overall layout qualityShould consider various constraints,multiple objectives,and new technologiesExplore effe

4、ctive&efficient machine learning(ML)techniques for placementAim to achieve better and faster solutionsGIEE,NTU4mixed-size objectsmacrostandard cellsThe EDA Lab1stRL-Based PlacementMirhoseini et al.,Nature,2021Formulate as a sequential Markov decision process(MDP),:State and action spaces,transition

5、dynamics,and reward signal Reinforcement Learning(RL)-based methodApplies an edge-based graph neural network(GNN)to generate a low-dimensional vector representationTrains a neural network to predict rewards on placements of new netlistsGroups millions of standard cells into few thousand clusters by

6、hMetis min-cut partitioningDiscretizes a few thousand grid cells and place macros and standard cell clusters onto the centers of the grid cellsGIEE,NTU5A.Mirhoseini et al.,“A graph placement methodology for fast chip design,”Nature,2021The EDA LabState-of-the-Art ML-Based Placement WorksGraphPlanner