A Predictive Thermal Model for Multiprocessor System-on-Chip

Recently, the increase of the transistor density in Multiprocessor system-on-chips (MPSoCs) and the constant rise of the operating frequency of the processor result in high on-chip temperature. The stability and reliability of MPSoCs inevitably have been seriously affected. Most thermal managements need regional temperature sensing to provide judgement, so temperature prediction adopting the thermal resistance and thermal capacitance (Thermal RC) model becomes an ideal solution to obtain the regional temperature conveniently and accurately. In this paper, we proposed a predictive thermal model based on the Thermal RC model combined with second derivative, which can increase the prediction time length to reduce the number of times that the module is invoked, and achieve the goal of reducing extra overhead. The experiment results demonstrated that, with the same margin of prediction error which is 0.6% (about 0.7 degrees C), the proposed predictive thermal model can increase the prediction length from is to 2.6s. Even during the period of Os-is, the maximum deviation of the relative prediction error between the proposed model and the contrastive model is 0.13% (about 0.16 degrees C).