LMS-based adaptive temperature prediction scheme for proactive thermal-aware three-dimensional Network-on-Chip systems

The three-dimensional Network-on-Chip (3D NoC) has been proposed to solve the complex on-chip communication issues. Because of the die-stacking architecture, the thermal problem becomes more severe than in 2D NoC. To simultaneously consider the thermal safety and system performance, proactive thermal management (PDTM) has been proved as an efficient way to control the system temperature against overheat. Based on the information of predictive temperature, the PDTM can early control the system temperature. To predict the future temperature, adopting the Thermal Resistance and Capacitance (Thermal RC) model is a popular way to derive the thermal prediction scheme. However, the Thermal RC value is sensitive to temperature changes, which affect the accuracy of the future temperature estimation. Therefore, the current proactive thermal-aware NoC system still suffers from large performance impact because of imprecise future temperature estimation. In this paper, we propose an LMS-based adaptive thermal prediction (LMS-ATP) model, which can adaptively adjust the involved Thermal RC values for future temperature estimation. The experimental results show that the proposed LMS-ATP model can improve the precision of future temperature estimation by 72.96%. In addition, the system throughput can be enhanced by around 0.77% to 47.96%.