LSTM-based Temperature Prediction and Hotspot Tracking for Thermal-aware 3D NoC System

As the increasingly large energy consumption and diverse workload distribution in many-core systems, thermal problems, especially hotspots, have become a significant restriction for the performance of the Network- on-Chip system. Therefore, several proactive thermal management methods based on thermal prediction data were proposed. However, previous temperature prediction methods did not consider the impact of hotspots, and their prediction accuracy is not satisfying. In this paper, we propose a Long Short-Term Memory (LSTM)-based temperature prediction and hotspot tracking model in a thermal-aware 3D NoC system. As the experiment indicates, the Mean-Square-Error (MSE) of the 6-step temperature prediction of the proposed model is 0.411 degrees C and the response time for tracking hotspot transfer is less than 0.075 ms.