Real-Time Temperature Prediction for Large-Scale Multi-Core Chips Based on Graph Convolutional Neural Networks

Abstract

The real-time temperature prediction of chips is a critical issue in the semiconductor field. As chip designs evolve towards 3D and high integration, traditional analytical methods such as finite element software and HotSpot face bottlenecks such as high difficulty in modeling, costly computation, and slow inference speeds when dealing with large-scale, multi-hotspot chip thermal analysis. To address these challenges, this paper proposes a real-time temperature prediction model for multi-core chips based on Graph Convolutional Neural Networks (GCNs) that includes the following specific steps: First, the multi-core chip and its temperature power information are represented by a graph according to the physical pattern of heat transfer; Second, three strategies—full connection, setting a truncation radius, and clustering—are proposed to construct the adjacency matrix of the graph, thus supporting the model to balance between computational complexity and accuracy; Third, the GCN model is improved by assigning learnable weights to the adjacency matrix, thereby enhancing its representational power for the temperature distribution of multiple cores. Experimental results show that, under different node numbers and distributions, our proposed method can control the Mean Squared Error (MSE) error of temperature prediction within 0.5, while the single inference time is within 2 ms, which is at least an order of magnitude faster than traditional methods such as HotSpot, meeting the requirements for real-time prediction.