This week marks several major milestones for Vinci. On the same day we emerged from stealth and announced our Series A funding, our technical team, including our co-founder & CEO Hardik Kabaria, presented Vinci’s first public research paper at IEEE EPTC 2025 in Singapore, titled “Thermal Sensitivity Analysis of 3D IC Face-to-Back Stacking Using Foundation Models for Physics.” It was co-authored by 9 Vinci employees (nearly a third of our entire team!), making it a substantial collaborative effort.
EPTC is one of the highest signal conferences for advanced packaging and 3D integration in the Asia Pacific region. It brings together the engineers and researchers who are defining the next generation of semiconductor technology. Launching Vinci to the public while debuting our initial technical results on this stage was a huge moment for our team.
Our paper demonstrates how Vinci’s foundation model for physics can run accurate, high-speed, full-package thermal sensitivity studies for advanced 3D stacks at manufacturing resolution.
Vinci simulated peak and average temperature responses to material changes across an entire 3D IC stack. The study identified the thermal interface material as the dominant design lever, with BEOL dielectrics secondary and hybrid bonding layers largely neutral for thermal mitigation priorities. This combination of speed, fidelity, and automated characterization shows how Vinci can change simulation workflows and design optimization for advanced packaging.
Key results
- Full fidelity at layout scale
Vinci ingested native industry standard layout files (OASIS, GDS, IPC-2581) and automatically handled meshing and mesh convergence, resolving layout features smaller than 7 nm in BEOL. - Manufacturing scale physics
The study covers a 3D stacked package with ten layers, from nanometer features in BEOL up through the full package. Vinci uses homogenization along with full-package analysis to capture realistic behavior. - AI speed with solver grade accuracy
Vinci Thermal ran 9,101 simulations with an average of 218 million elements each in 26 hours total runtime (an average of 10.29 seconds each). The BEOL material characterization alone includes 432 solves on 300 million degree of freedom grids. It completes in 52 minutes on eight AMD Instinct MI300X GPUs. - Verified against commercial tools
Vinci matched the performance of a legacy solver on full package thermal results using averaged properties derived from Vinci’s own material characterization.
Our EPTC paper is one of the first public demonstrations of Vinci’s ability to operate at the scale and fidelity that semiconductor companies require. It validates our core thesis that AI, when rooted in physics, can accelerate hardware simulation and unlock new design space while preserving accuracy.
For our team, it also represents the beginning of a long and ambitious path. We are grateful to the researchers, reviewers, and industry partners who have encouraged this work. The response to our launch has been energizing, and we are committed to supporting the engineering community with tools that meaningfully improve how complex hardware is developed.
If you would like to review the full study, you can download the paper here.
