Continuous Physics Reasoning: Physics as Always-On Design Infrastructure

AI workloads are redefining thermal simulation requirements for advanced packaging. 2.5D HBM integration, chiplet-based accelerators, and 3D architectures create complex steady-state and transient thermal challenges that traditional workflows struggle to address.

This presentation introduces AI-accelerated thermal simulation that resolves trillions of degrees of freedom at manufacturing resolution in seconds, without manual meshing/simplification, validated against commercial FEA solvers. Two worked examples: 1) In a thermal co-design study of a ~1kW 2.5D HBM package, an 8-variable, 1,600-run DOE (within 0.8°C of Flotherm) shows the package is cooling-limited: only 1 of 1,600 designs meets the ≤85°C DRAM-core limit, collapsing a ~39-day sweep to 2 days. 2) In transient analysis under AI mission profiles, per-die junction-temperature histories resolve a 135m°C excursion inside a 110°C package, capturing cross-die coupling and cycle-to-cycle ratcheting for reliability and throttling-margin design.

These capabilities make thermal engineering an integral part of early package co-design.