Evaluating System Identification Methods for Predicting Thermal Dissipation of Heterogeneous SoCs

In this paper we evaluate the use of system identification methods to build a thermal prediction model heterogeneous SoC platforms that can be used quickly predict temperature different configurations without need hardware. Specifically, focus on modeling approaches based clock frequency and utilization percentage each core. We investigate three with respect their accuracy: linear state-space approach using polynomial regressors, NARX neural network recurrent configured in an FIR structure. Odroid-XU4 board featuring Exynos 5422 SoC. The results show regressors significantly outperformed other two models when trained 1 hour 6 hours data.