DeepOHeat: Operator Learning-based Ultra-fast Thermal Simulation in 3D-IC Design [Paper]
Thermal-aware floorplanning is particularly important in designing 3D integrated circuits. To reduce turnaround time in design optimization loops and avoid collecting data in substantial costs, we propose DeepOHeat with the following features:
- For the first time, DeepOHeat solves a family of heat equations under various 3D-IC configurations in an end-to-end, data-free scheme.
- We encode 3D-IC configurations using a modular approach: stacked cuboidal geometry, boundary conditions, 2d/3d power maps and full chip material conductivity.
- We achieve satisfactory accuracy when benchmarking DeepOHeat against Cadence Celsius Thermal Solver.
Figure 1. DeepOHeat framework.
Python 3.8.16 was used to develop this repo, to install dependencies
pip install -r requirements.txt
To reproduce examples in our paper or directly use pretrained checkpoints, please refer to 2d_power_map and multi_htc_bc.
- In particular, prototype.py directly runs inference on the 10 test cases we showcased in the paper, at Section V.A.
- Side by side comparison against Cadence Celsius is in the benchmark folder of each experiment's logging directory, if provided.
The single_case_pinns directory contains examples solving single instances of various 3D-IC configurations.
- complex_geometry showcases a geometry of two stacked cuboids with a volumetric power defined in the middle of the bottom cuboid.
- quarter_power and uniform_power showcase 2d surface power on top with different sizes.
- volumetric_power showcases a single cuboidal geometry with a volumetric power.
Figure 2. Benchmark DeepOHeat against Cadence Celsius. Top -> bottom: Celsius, DeepOHeat, point-wise error.
If you find this repo useful, please cite our paper.
@inproceedings{liu2023deepoheat,
title={DeepOHeat: operator learning-based ultra-fast thermal simulation in 3D-IC design},
author={Liu, Ziyue and Li, Yixing and Hu, Jing and Yu, Xinling and Shiau, Shinyu and Ai, Xin and Zeng, Zhiyu and Zhang, Zheng},
booktitle={2023 60th ACM/IEEE Design Automation Conference (DAC)},
pages={1--6},
year={2023},
organization={IEEE}
}
DeepOHeat is provided under MIT License, please see LICENSE for full license text.
If you have any question regarding the code or the paper, please contact ziyueliu@ucsb.edu.
This is a joint work between Zhang's group at UCSB and Celsius R&D team at Cadence Design Systems.