Performance-per-watt optimal control logic in liquid cooling solution for high performance machine-learning computing

摘要

An electronic rack includes an array of server blades arranged in a stack. Each server blade contains one or more servers and each server includes one or more processors to provide data processing services. The electronic rack further includes a coolant distribution unit (CDU) and a rack management unit (RMU). The CDU is configured to supply cooling liquid to the processors and to receive the cooling liquid carrying the heat exchanged from the processors. The CDU includes a liquid pump to pump the cooling liquid and a pump controller to control a pump speed of the liquid pump. The RMU is configured to manage the operations of the components within the electronic rack such as CDU, etc. The RMU includes control logic to determine an optimal pump speed of the liquid pump by optimizing an objective function based on processor power consumption of the processors, cooling power consumption of the CDU, and a benchmark performance measurement of the data processing services and to control the pump speed of the liquid pump based on the optimal pump speed.