Modeling and Deep Explicit Model Predictive Control for Server Processor Direct Liquid Cooling

Direct liquid cooling greatly improves the capacity and efficiency for high thermal designed power (TDP) processors. Advanced micro-structured cooling devices and loop system has been designed. Given rapid changing working load condition and complex thermal dynamics of the system, advanced control method is needed to balance cooling power and energy consumption in order to improve operating efficiency. This paper presents the design and implementation of an explicit model predictive controller based on neural network for the liquid cooling system. This neural network explicitly represents the decisions in terms of observations. Loss function for training the neural network is derived from the paradigm of model predictive control, where the model is obtained from physical-empirical-hybrid simulations. The training data is sampled uniformly from the working point of the system, and training process is done on cloud with model. The well-trained model is then implemented as controller in a local experimental platform. Experiments on cooling efficiency performance under randomly varying loads show that this controller has better temperature control and reduced pump energy consumption comparing to traditional PID controller. This result achieves progress on: 1) characterizing thermodynamics of the processor liquid cooling system by simplified and approximated model; 2) building experimental platform to support testing of liquid cooling control algorithms; 3) developing an explicit form of model predictive controller approximated by deep neural network (DEMPC); and 4) demonstrating the performance of the DEMPC over the traditional proportional controller.