(Invited, Digital Presentation) ML-Based Fast on-Chip Transient Thermal Simulation for Heterogeneous 2.5D/3D IC Designs

Abstract

Large 3DIC designs with multiple chips requireseveral iterations of transient thermal analysis particularly forfine-grain on-chip dynamic thermal management. This requiresa fast thermal analysis technology as opposed to traditionalCFD/FEA based methods which have severe runtime/capacitylimitations for large chips (e.g., 2cmx2cm) in 3DIC whilegenerating fine grained (e.g., 10umx10um) transient thermalresponse. The fast transient thermal analysis is based on the ideaof combining the global, intermediate, and local transientresponse curves generated from an ML-predictor. The local,intermediate, and global transient response curves are scaledbased on the far-field and near-field transient decay surfacecomponents respectively, computed using the trained ML decaysurface predictor, followed by linear superposition of the curvesfor each power value in the transient power profile to generatethe effective transient response curve. The runtime forgenerating thermal results for a large chip is in the order ofminutes, compared to several hours/days while using CFD/FEAbased tools with good accuracy correlation. The fast transientthermal solver is implemented on distributed ML computingplatform for parallel computation of transient thermal inferencemodel.