Real-time thermal management of 3D multi-core system with fine-grained cooling control

This paper presents a cyber-physical (real-time sense-predict-adjust) thermal management for 3D multi-core system by micro-fluidic cooling. Auto Regressive (AR) model is used to predict future workload and the prediction is furnished by Kalman filtering to get rid of noises due to system variation. A thermal model is developed to sense thermal behaviour of the 3D system including micro-fluidic channels and estimate future thermal demand. The use of fine-grained, or non-uniform, flow-rate control with channel clustering is then incorporated to adjust flow-rates based on the predicted future thermal demand in a real-time fashion. Experiment results show that under this cyber-physical scheme, the temperature of 3D multi-core cache-processor system is well maintained below threshold and a more even temperature distribution is achieved with lower fluid-pump power or total flow-rate overhead. For example, the total flow-rate has a significant reduction of 72.1% under the fine-grained flow rate control.