ThermalGCN

ThermalGCN is a fast graph convolutional networks(GCN)-based method for thermal simulation of chiplet-based systems.

• Use global information (total power) as input,
• Apply the skip connection in graph convolution network,
• Integrate PNA network into the model,
• Use edge based attention network to represent the connection effect.

Installation

ThermalGCN requires Pytorch and DGL to be installed as backend.

Instructions

• Random chiplet layout generation:

  cd ./dataset/

  python Generate.py

• Obtain dataset:

  create a folder named "data".

  python run.py to run hotspot and generate dataset which is stored into "./dataset/data".

  python data_preprocess.py to normalize the data.

• Training GCN:

  python GCNPNAGAT.py

Publications

L. Chen, W. Jin and S. X.-D. Tan, "Fast Thermal Analysis for Chiplet Design based on Graph Convolution Networks," 2022 27th Asia and South Pacific Design Automation Conference (ASP-DAC), 2022, pp. 485-492..

The Team

ThermalGCN was originally developed by Liang Chen and Wentian Jin at VSCLAB under the supervision of Prof. Sheldon Tan.

ThermalGCN is currently maintained by Liang Chen.