基于图像到图像机器学习的稳健技术可迁移静态IR压降预测.pdf

1、The EDA LaboratoryRobust Technology-Transferable Static IR Drop Prediction Based on Image-to-Image Machine LearningThe Electronic Design Automation LabNational Taiwan UniversityC.-C.Lan,C.-C.Su,Y.-H.Lu,and Yao-Wen Chang 1Outline2 Introduction Methodology Experimental Results ConclusionStatic IR Drop

2、Voltage drop between the power supply and circuit instances under steady-state conditionsCaused by the resistance of metals and vias in the power delivery network(PDN)Calculation of static IR drop invloves solving linear equations,where each equation corresponds to a node inside the PDNBecome extrem

3、ely time-consuming for large designs3Machine Learning-based Prediction WorksCell-based prediction Pao et al.,DATE20,Ho et al.,ICCAD19,and Kundu et al.,VLSID22Predict IR drop for a single cell per inferenceLack design transferabilityImage-to-image prediction Chhabria et al.,ASPDAC21Predict IR drop fo

4、r the whole circuit per inference with U-Net Ronneberger et al.,MICCAI15Achieve design-transferabilityLack technology-transferability4The U-Net architectureMotivation5Previous works face challenges in real world application due toLack of technology-transferabilityImbalance training data across diffe

5、rent technologiesA prediction methodology that can transfer knowledge from one technology to another is neededOverview6We proposed the first technology-transferable prediction methodology for static IR dropTechnology 1Technology 2Target TechnologyModelFine-tuningPrediction GoldenModelPre-training an

6、d Fine-tuning7Data AugmentationPre-training DataPre-trained ModelData ProcessingModel TrainingPre-training PhaseFine-tuning PhaseDifferent Technology CircuitsSimple Data AugmentationFine-tuning DataFine-tuned ModelData ProcessingModel Fine-tuningTarget Technology CircuitsTesting PhaseIR Drop Predict