DNN-based Fast Static On-chip Thermal Solver

TLDR

Accurate on‑chip temperature prediction is critical for performance and reliability in 5G, automotive, and AI chip‑package systems, yet finite element or CFD methods are time‑consuming and thermal gradients influence electromigration, aging, and design constraints. The study aims to deliver a fast yet accurate on‑chip thermal solution for chip‑package and 3DIC designs. This is achieved with a data‑driven DNN‑based thermal solver that is 100–1000× faster than traditional FEM while preserving the same level of accuracy.

Abstract

Accurate prediction of on-chip temperature distribution becomes important for the performance and reliability of upcoming 5G, automotive, and AI chip-package-systems. In particular, a large thermal gradient (the temperature variation across a chip) accelerates electromigration and aging, and also impacts design performance and power. Furthermore, there are usually Tmax (maximum temperature) constraints on junctions of a chip, skin temperature concerns for mobile devices or wearables, and important placement considerations of on-chip thermal sensors for use in dynamic voltage and frequency scaling. However, obtaining an accurate and detailed thermal gradient on-chip is very time-consuming using the finite element method (FEM) or computational fluid dynamics (CFD) technology. Furthermore, there are many different functional scenarios for various applications that users need to identify possible Tmax locations on-chip. Therefore, there is an urgent need in the industry to provide a fast, yet accurate on-chip thermal solution in a chip-package-system or more complicated 3DIC design, which may include multiple chips. This paper proposes a method to use a data-driven DNN-based thermal solver that can be 100-1000x faster depending on the size of the chip compared to traditional FEM-based thermal solvers with the same level of accuracy.