Transient Thermal Analysis of Chiplet-Based Systems with Dual-Channel Message Passing Neural Networks

This paper presents a novel transient thermal analysis method for chiplet-based systems using the dual-channel message passing neural networks (MPNNs). Firstly, the package information is transformed into graph-structured data by using the finite difference approximation. Secondly, the temperature and power features are extracted through the dual-channel MPNNs to obtain the temporal gradient of temperature. Finally, the augmented neural ordinary differential equations are employed to learn the continuous-time dynamics of the heat transfer equation. The experimental results demonstrate that the model can accurately estimate the transient thermal maps with the mean absolute error of 0.272 K and the root mean squared error of 0.387 K. It is shown that the model can predict thermal maps at arbitrary time points, exhibiting good transferable capability.