Dynamic Thermal Management by Controlling CPU Frequency and Fan Speed

Microprocessor designers use the design of multicore systems on a chip to increase their computing power. Adding the number of cores leads to an increase in the power density, followed by enhancement of temperature. Reactive and proactive approaches are two sets of the methods for managing the temperature. Unlike the reactive methods that act based on threshold temperature, proactive approaches utilize a thermal prediction model in thermal management. In this paper, two multilayer perceptron neural networks has been used for thermal prediction and temperature control. An appropriate dataset is provided for training each model. This dataset consists of some features that are read by sensors and measurement tools and new features that are produced by proposed processes. In this regard, historical features are suggested for thermal model. Proper features of thermal model are selected by using feature selection based on mutual information. The temperature is controlled by setting the processor frequency and fan speed. The features of control model are selected by non-dominated sorting genetic algorithm. The error of thermal model for different time distances is about 0.5 °C. The thermal control model has respectively 2% and 0.6% errors in determining the processor frequency and fan speed.