Dynamic Thermal Management for 3-D ICs with Time-Dependent Power Map Using Microchannel Cooling and Machine Learning

It is agreed that air-cooled heat sink (ACHS) would become incapable of 3-D integrated circuits (ICs). A switch from ACHS to a microfluidic heat sink (MFHS) is believed to be a promising solution. Tier-specific MFHS, where the flow rate of each tier can be controlled independently, has been further proposed in consideration of the power consumptions of pumps. However, these works are generally based on a steady power map, which in reality is mostly time-dependent. In this paper, a machine learning (ML)-based control method, combining the Bayesian optimization (BO) and the artificial neural network (ANN), is applied to 3-D ICs with the tier-specific MFHS, considering a time-dependent power map. BO is first applied because it has been demonstrated to outperform other state-of-the-art black-box optimization techniques due to its quick converge. However, as more and more data are acquired when the system keeps working, its computational time increases sharply due to the increasing calculation complexity, which cannot be accepted as we aim for dynamic thermal management. Therefore, ANN is then applied. With the online learning method, its calculation complexity remains constant as more data are acquired. Because of this, its time consumption remains small as the system keeps working. Results of the flow rates and the temperatures are finally presented, which prove that with the ML-based control method, power consumptions of the pumps are intelligently saved while, at the same time, the temperature constraints are met.