Server Level Impacts on CPU Cooling Capability in Single-Phase Immersion

The continued demand for increased compute performance has been driving components, such as CPUs and GPUs, and servers to higher power and power densities, to the point where air-based cooling technologies are reaching their limits of cooling capability at reasonable operating expenses/costs (OpEx). High performance data center servers are therefore steering towards more cost effective and efficient liquid cooling technologies such as cold plates and immersion. Single-phase immersion provides a low cost, efficient, sustainable, easily adaptable, and holistic cooling technology for the entire server/IT equipment. Immersion can also be neutral to ambient conditions and air contaminants, making it a preferred cooling technology for edge applications.In immersion, the IT equipment can be immersed vertically in a tank filled with a dielectric fluid such as Polyalphaolefin (PAO). As the IT equipment operates, heat generated by the powered components creates a natural convection flow that drives the fluid through the server (i.e. hot liquid rising), thus cooling the components. This cooling may be further enhanced by using optimized immersion heat sinks for high-power components and/or forcing flow through the IT equipment if needed. The heated exiting fluid is then cooled by the facility water via a Coolant Distribution Unit (CDU) and pumped back into the tank and IT equipment. While tanks in the industry today provide a wide range of fluid manifolds and pump capacities, there is very limited understanding of the actual flow rate boundary conditions for a server in a tank. This poses a challenge to design thermal solutions for the high-power components and to estimate the cooling capability for the IT equipment in a tank. Hence, understanding the specific impacts of different IT components, configurations and boundary conditions on the thermal performance is essential to design efficient thermal solutions and to predict system cooling capabilities.In this paper, an Intel spreadcore server (code name: Coyote Pass, M50CYP) is used to investigate the cooling capability of the 2 CPUs in the 1U server. The two CPUs are 3rd Generation Intel® Xeon® Processor Scalable Family (code name: Ice Lake) CPUs and an immersion heat sink optimized for natural convection is installed on each one. The CPU thermal cooling capability is evaluated under various workloads while running Intel Power and Thermal Analysis Tool (PTAT) at maximum Thermal Design Power (TDP) of 205 W.The impact on the CPU cooling capability by powering on other server components is investigated in natural convection. The results show that the CPU thermal resistance values improve when surrounding components (storage drives and memory modules) are powered on.The impact on the CPU cooling capability by forcing a flow rate through the server is also investigated. The flow rate ranging between 0.9 lpm to 10 lpm. The results show that the CPU thermal resistance values deteriorate if the server flow rate is below the flow rate naturally generated by natural convection. The thermal resistance values improve as the flow rate increases, and at a flow rate of 10 lpm through the server a ~30% improvement is determined.