How Full Liquid‑Cooling Servers Achieve Near‑100% Heat Capture: Design, Flow, and Component Insights

System Composition and Pipeline Layout

The full liquid‑cooling server is built on the Inspur 2U four‑node high‑density compute platform i24. Each node supports two Intel 5th‑gen Xeon scalable processors, 16 DDR5 DIMMs, one PCIe expansion card, and one OCP 3.0 network card, with up to eight SSDs. Major heat sources are CPU, memory, I/O boards, local SSDs, and the power supply.

The system uses a serial flow path: low‑power components are cooled first, and the coolant then passes to higher‑power devices, as shown in the diagrams.

Flow Design and Rate Calculation

To ensure long‑term reliability of secondary‑side piping, the return water temperature must stay below 65 °C. The design targets a copper cold plate with PG25 coolant. Using the formula Q_min = P_sys / (ρ·C·ΔT), the minimum flow per node is about 1.3 LPM, achieving roughly 95 % heat removal by direct liquid contact and the remaining ~5 % by a rear‑mounted air‑liquid heat exchanger.

Key Cold‑Plate Component Designs

1) CPU Cold Plate

The CPU cold plate is a reference design for Intel 5th‑gen Xeon processors, optimized for thermal performance, structural integrity, yield, cost, and material compatibility. It consists of an aluminum bracket, the cold plate itself, and a connector.

2) Memory Liquid‑Cooling

The memory solution uses a novel “sleeper‑rail” heat‑sink that combines traditional air cooling with a cold plate. Heat from the DIMMs is transferred via built‑in heat pipes to the heat sink, then through a thermal pad to the liquid‑cooled cold plate. The memory module can be assembled as a maintenance unit, with optional screw or tool‑free fixation, and can also provide cooling for other board components.

Easy maintenance: memory modules are serviced like air‑cooled modules without removing heat‑sink hardware.

High compatibility: works with memory spacing from 7.5 mm upward.

Cost‑effective: heat‑sink material and design can be selected based on memory power, supporting >30 W per module.

Simple manufacturing: no internal liquid tubing, allowing use of standard fan‑cooled heat‑sink processes.

Reliability: avoids damage to DIMM chips or thermal pads during assembly and supports repeated hot‑swap operations.

3) SSD Liquid‑Cooling

The SSD solution integrates a heat‑pipe‑based heat sink with the SSD module, a cold plate, a locking mechanism, and a bracket. The lock provides pre‑load to maintain long‑term contact between SSD and cold plate, while the bracket allows drawer‑style installation in the server chassis.

Supports >30 hot‑swap cycles without power loss.

No shear stress on thermal interface material during installation.

Low manufacturing complexity, using existing fan‑cooled SSD processes.

Water‑free design enables multiple SSDs on a single cold plate, reducing leak risk.

Adaptable to various SSD thicknesses and quantities.

4) PCIe/OCP Card Liquid‑Cooling

PCIe: A custom liquid‑cooled heat‑sink module contacts the PCIe card’s main chips and optical module. Heat pipes transfer heat to the module, which then contacts the I/O cold plate via thermal interface material. The design includes a QSFP heat‑sink clamp with appropriate elasticity to ensure stable contact.

OCP 3.0: Similar to PCIe, the OCP card uses a dedicated liquid‑cooled heat‑sink. The module is mounted with a spring‑screw lock to maintain long‑term contact with the I/O cold plate.

5) I/O Cold Plate

The I/O cold plate serves multiple functions: cooling the motherboard I/O area, and providing a thermal interface for liquid‑cooled PCIe and OCP cards. It consists of an aluminum alloy body and copper tubing for coolant flow, designed according to board layout and component heat requirements.

6) Power‑Supply (PSU) Cold Plate

The PSU liquid‑cooling solution adds an external air‑liquid heat exchanger to a standard fan‑cooled power supply, removing hot exhaust air without redesigning the PSU. The rear‑mounted heat exchanger is a multi‑layer structure with overlapping flow channels and fins, sized to fit within rack constraints while balancing cooling performance, weight, and cost.

This design avoids the need for a dedicated liquid‑cooled PSU, shortening development cycles and reducing cost by over 60 % compared with custom solutions. For rack‑level deployment, a centralized air‑liquid heat exchanger can be installed at the rack’s front and rear doors, providing up to 8 kW cooling capacity for more than 150 nodes while maintaining a closed‑loop, low‑impact thermal environment.

Overall, the full liquid‑cooled server removes approximately 95 % of heat directly via cold plates, with the remaining <5 % handled by the rear air‑liquid heat exchanger, delivering near‑100 % heat capture efficiency.