Open Benchmark Evaluates AI Thermal Models for 2.5D and 3D ICs
A cross-institutional research effort from the University of Technology Sydney (UTS), TU Munich, and ShanghaiTech University delivers a public benchmark for assessing machine-learning-based thermal prediction in advanced packaging.
Why Thermal Modeling Matters for 2.5D and 3D ICs
As the semiconductor industry scales beyond conventional monolithic designs, 2.5D interposer-based and 3D stacked architectures have become mainstream vehicles for heterogeneous integration. By 2026, chiplets assembled on silicon interposers and vertically stacked dies are commonplace across high-performance computing, AI accelerators, and advanced mobile SoCs. The payoff β higher bandwidth, shorter interconnect, and mixed-process integration β comes with a persistent engineering challenge: heat.
Heat density in stacked and side-by-side die configurations is far less forgiving than in planar chips. Thermal hotspots can degrade performance, accelerate electromigration, and shorten device lifetime. Accurate thermal maps are therefore essential during design-space exploration, but full-physics finite-element or finite-volume simulation is computationally prohibitive when swept across thousands of layout variants.
The Rise of AI Surrogates for Thermal Simulation
Machine learning surrogates have emerged as an attractive alternative. Trained on a modest number of high-fidelity simulations, a neural network can approximate temperature fields orders of magnitude faster, enabling rapid design iteration. The catch: without a shared, reproducible evaluation framework, claims of accuracy and speed across published models are difficult to compare.
That gap is precisely what the UTSβTU MunichβShanghaiTech collaboration addresses.
What the Benchmark Provides
According to the researchers, the benchmark establishes a common testbed for AI-driven thermal models targeting 2.5D and 3D ICs. Key elements include:
- Standardized test cases drawn from representative 2.5D and 3D configurations, ensuring models are compared on identical geometries and power maps.
- Reference solutions generated by established numerical solvers, providing a ground truth against which surrogate predictions can be scored.
- Consistent metrics for accuracy (e.g., temperature error distributions, hotspot localization) and computational cost.
- Open availability, allowing academic and industrial groups to reproduce results and submit new models for comparison.
Why an Open Benchmark Is Significant
The absence of agreed-upon evaluation standards has been a quiet drag on progress. Different papers use different geometries, power profiles, and error metrics, making it nearly impossible to judge whether one architecture genuinely outperforms another. An open benchmark changes the incentive structure: researchers can build on each other's results rather than re-deriving baselines, and industry practitioners gain a clearer signal about which approaches are production-ready.
For 2.5D and 3D integration specifically, the benchmark is timely. Chiplet-based design is now a default strategy for many advanced nodes, and thermal co-design has become a first-order concern rather than a post-layout afterthought. Tools that can deliver fast, reliable thermal estimates early in the flow have direct commercial value.
Looking Ahead
As heterogeneous integration continues to deepen β with taller stacks, finer bump pitches, and more aggressive power densities β the demand for fast, accurate thermal surrogates will only grow. Open benchmarks like this one provide the shared infrastructure needed to separate genuine advances from incremental tuning, and to accelerate the translation of ML research into practical EDA tooling.
The collaboration between UTS, TU Munich, and ShanghaiTech also underscores a broader trend: thermal management for advanced packaging is now a global research priority, drawing on expertise from across continents and disciplines.