Transfer learning of convolutional neural network model for thermal estimation of multichip modules

Abstract

This paper proposes a transfer learning approach to reduce the dependence of the neural network model on dataset size for multi-chip modules (MCMs) thermal estimation. The dataset that resembles the target task is used to pre-train a convolutional neural network model; then, it is fine-tuned using a much smaller dataset belonging to the target task. This process enables the model to adapt and refine its thermal estimation capabilities for the particular MCM under consideration with sparse data. Results show that the model effectively utilizes the geometric shapes and power details of MCMs to precisely predict steady-state temperature fields and junction temperatures. In a complex five-chip configuration, the fine-tuned model achieves a prediction error of only 1.17 % and accelerates inference speed by three orders of magnitude compared to traditional simulations. Additionally, the transfer learning model requires only 1/8 the dataset size and less than 1/3 the training time compared to a model with random initial parameters to achieve similar accuracy. The results indicate that this model can significantly enhance the utilization efficiency of historically accumulated MCM heat transfer datasets for new demands such as real-time thermal estimation and fast optimization of chip configuration.

Keywords

• Transfer learning
• Multi-chip modules
• Thermal design
• Temperature reconstruction