Multimodal Learning-Based Thermal Solver for 3DICs with Arbitrary Chip Dimensions and Power Distributions

Thermal management is critical for 3DIC design because stacking dies vertically increases power density and peak temperature and creates thermal hotspots, which can significantly impair transistor reliability, degrade performance, and accelerate device aging. Conventional static and transient thermal analysis approaches use finite element method (FEM) and computational fluid dynamics (CFD) to provide detailed and accurate thermal profiles of chips. However, they are computationally intensive and impractical for today’s large-scale 3DIC designs. Recently, machine learning (ML) is used to accelerate thermal predictions, but a sufficiently large training dataset is needed to re-train or fine-tune models for chips of different dimensions and/or power distributions. In this paper, we propose a multimodal learning-based framework that can effectively learn from limited training data and accurately predict the thermal behavior of diverse chip configurations. The model is trained only once with data of fixed size, which is much smaller than that required by the current ML methods. Compared to a commercial electrothermal simulator, our approach achieves a 150× speedup of runtime and requires only 0.023% of the memory. On chips of different die sizes, our model’s prediction error is within 2.63% of the thermal prediction of a commercial simulator, and less than 3.72% for real-life power patterns.