As chip designs move to 2.5D and 3D stacking, heat dissipation becomes harder to model. Traditional simulations are slow, so engineers are turning to machine learning for faster thermal predictions. But without a common yardstick, it is difficult to know which AI models actually work across different package geometries and materials.
The new benchmark, called IC-ThermBench, tries to fill that gap. Created by researchers at the University of Technology Sydney, ShanghaiTech University, and the Technical University of Munich, it is described as an open and progressive test suite. That means it is publicly available and designed to grow in difficulty, pushing AI models to generalize beyond the specific cases they were trained on.
The paper notes that existing thermal learning approaches often fail when applied to new chip layouts or thermal boundary conditions. IC-ThermBench aims to expose those weaknesses by offering a structured, reproducible way to measure performance. The researchers hope this will help the field converge on more robust models for real-world 2.5D and 3D IC design.