Advanced Cold Plate Liquid Cooling Solution for Hyper-scale Data Center Application

With the rapid development of 5G, artificial intelligence(AI), big data and cloud computing, high performance and density data centers have become the fundamental infrastructure to address the continuous demand for cost-efficiency and energy-efficiency application. In order to cope with the performance and density request, the power consumption of CPU and GPU chips has significantly increased year by year, resulting in huge thermal design challenges with traditional air-cooling solution. For example, Thermal Design Power (TDP) of 3
^rd
Gen Intel Xeon Scalable processors (codenamed Ice Lake) is up to 270W and TDP of Nvidia A100 GPU is up to 500W. This is almost beyond the cooling capability of traditional air-cooling solution under current thermal boundary condition due to limitation of heatsink size and air flow. In addition, Power Usage Effectiveness (PUE) is playing an increasingly key role on newly built data centers and has become mandatory requirement in government regulations. Under these circumstances, advanced cooling solution is urgently requested. Liquid cooling solution is an efficient solution to tackle the increasing power density in the data centers and becomes increasingly popular in modern data centers due to its excellent cooling capability, energy efficiency and PUE. This paper introduces the engineering practice of cold plate liquid cooling solution for hyper-scale data center application adopted by Volcano Engine, the cloud service platform under ByteDance.COM, covering the architecture and design from facility, server to components. This paper also includes the design and optimization of cold plate as well as other ingredients as coolant distribution unit(CDU), manifold, cold plate, tube, liquid leakage detection methodology. The thermal performance and power consumption are also analyzed and compared with air cooled system. The quality and reliability results are summarized based on more than six months of functional testing with over one thousand of pilot deployment servers powered with Ice Lake processors under cold plate liquid cooling solution in data center.