An Accurate Deep Learning-Based Thermal Reconstruction Technique for Microprocessors Using Embedded Sensors

This paper proposes a deep learning-based thermal reconstruction technique to accurately recover a microprocessor's full thermal maps with a limited number of sensors. A neural network model derived from U-Net is improved to infer comprehensive thermal profiles using limited sensor readings. An integrated thermal simulation framework combining Gem5, McPAt and HotSpot generates training data, and evaluates the technique on a microprocessor running real applications. Results demonstrate that the proposed technique can accurately reconstruct thermal maps and outperform prior methods.