T-Fusion：基于多保真度融合的 3D IC 热预测.pdf

1、1T-Fusion:Thermal Prediction of 3D ICs with Multi-fidelity FusionBingrui Zhang,Wei W.Xing*,Xin Zhao,and Yuquan Sun*Beihang University&The University of SheffieldASP-DAC 2025OutlinesBACKGROUND AND PREVIOUS WORKMETHODOLOGYEXPERIMENT RESULTSSUMMARYMotivation:Why care about Thermal ModellingMoores Law i

2、s slowing down 3D Integration offers a solution But brings thermal challenges:Higher power densityComplex heat pathsPerformance degradationNeed fast and accurate thermal analysis(a)Air-cooled heat dissipation (b)liquid-cooled heat dissipationTraditional Methods:numerical solversCOMSOL(1000+seconds)H

3、otspot(1450+seconds)MTA(200+seconds)ML-based Methods:learn the mappingDeep OHeat(100+hours training)ThermPINN(2D only)Therm-Transformer(large data needed)SOTA and limitationNeed fast and accurate thermal prediction with limited dataMulti-fidelity ApproachHigh-fidelity:Accurate but expensiveLow-fidel

4、ity:Fast but less accurate Fusion:Best of both worlds(a)Chip Section Diagram(b)Low Fidelity Mesh(c)High Fidelity MeshInsightsLeverage both Low-and High-Fidelity as data sourceKey requirementsLarge-scale temperature fieldsDifferent resolutions across different fidelitiesSpatial correlationTemporal co

5、nsistencySmall training datasetScalabilityChallenge:MethodologyT-Fusion ModelLow-fidelity modelTensor multiplicationResidual modelKey innovation:Tensor based cross fidelity transformationPreserve spatial relationshipResolution alignmentReducing computation from O(d3)to O(d2)Efficient TrainingL=Ll+Lr

6、Reduce memoryExperimental Results:BenchmarkTest Cases of 3D ICssingle-corequad-coreeight-core chipsInput:Power consumption for each function blockOutput:Spatial-temporal temperature profileExample Power Differences in Three 3D-ICsExperiment platformIntel Xeon CPU 2.40GHz NVIDIA A100 PCIe 40GB GPU 12