利用来自实际设计的有限数据进行静态 IR 压降预测.pdf

1、Static IR Drop Prediction with Limited Data from Real DesignsLizi Zhang,and Azadeh DavoodiPresented by Robert ViramontesDepartment of Electrical and Computer Engineering University of Wisconsin-MadisonOutline ML for IR-drop prediction Our methodology Simulation results Conclusion1ML for IR-drop pred

2、iction:Introduction Static IR drop analysis of power delivery network(PDN)is a crucial task in IC design Voltage drop is induced between the power pads and cells in the design High IR drops can severely impact the normal functionality of chips Traditional techniques require solving a system of equat

3、ions representing KVL and KCL and can take up to few hours using commercial tools 2ML for IR-drop prediction:prior work Earlier work for PDN analysis trade off accuracy for speed by utilizing techniques like spatial localityDAC11,preconditioned conjugate gradient ICCAD11.Recently,machine learning(ML

4、)-based techniques provide significantly faster and more accurate solutions Limited to incremental analysis VTS18,VLSID22 Applicable to specific designsASPDAC20 Not accurate enough ASPDAC213ML for IR-drop prediction:challenges1.Complexity of PDNs 3D network with up to 10 layers Hard to make accurate

5、 predictions2.Lack data from real chips for AI-based prediction Large amounts of training data are often required to produce accurate predictions4Methodology:overviewIR drop prediction with AttUNet5 Translate the original problem to image-to-image prediction Train our model using the two-step pretra

6、in-finetune strategy Make prediction and evaluateMethodology:image-based inputs&output6Methodology:image-based inputs&output The Power delivery networks can be modeled as a 3D grid of voltage sources,current sources,and resistances Wires are a network of resistances,the power pad(C4 bumps)are voltag