Online Machine Learning-based Temperature Prediction for Thermal-aware NoC System

The Network-on-Chip (NoC) has been proposed to solve the communication problem in multicore systems, but it usually suffers from the serious thermal problem due to high power density. To solve this problem, the proactive thermal management (PDTM) has been proven as an efficient way to prevent the system from overheating and mitigate the performance impact during the temperature control period. Based on the predicted temperature results, the PDTM controls the system temperature in advance to make the system temperature under the thermal limit. However, due to the thermal-coupling effect on the chip, it is hard to have a precise thermal prediction model, which makes the PDTM cannot control the system temperature efficiently. In this paper, we propose a lightweight thermal prediction model based on the machine learning method accompany with an online training algorithm that can adapt the hyperplane of the temperature behavior of NoC system during the runtime. The proposed model can adapt varying situations of the temperature behavior of NoC systems on the fly. Compared with the traditional thermal prediction model, the proposed approach can reduce 40.6-51.5% average error and 39.4%-54.2% maximum error.