What Is Direct-to-Chip Liquid Cooling?
Direct-to-chip liquid cooling is a thermal management method in which a liquid coolant is circulated through a cold plate mounted directly on the processor package — GPU, CPU, or high-power memory. The cold plate absorbs heat from the chip surface and transfers it to the coolant, which carries the heat away to a cooling distribution unit (CDU) or facility water circuit.
Unlike air cooling, which relies on airflow across heatsinks, direct-to-chip cooling can remove heat at much higher flux densities. This makes it the primary cooling method for AI accelerators and high-performance computing (HPC) processors that exceed the thermal limits of air cooling.
Why AI Servers Require Liquid Cooling
AI training and inference workloads are computationally intensive. Modern AI accelerators — GPU clusters used for large language model training, for example — generate heat at densities that air cooling cannot efficiently remove within the physical constraints of a 1U or 2U server chassis.
Air cooling requires large volumes of airflow and significant fan power to move heat from the chip to the room air. As chip thermal design power (TDP) increases, the airflow required grows, increasing fan noise, fan power consumption, and the risk of hot spots within the chassis. Direct-to-chip liquid cooling removes heat at the source with a compact cold plate, reducing the dependence on airflow and enabling higher rack densities.
Cooling Loop Architecture
A direct-to-chip cooling system consists of a primary loop (facility water or chilled water) and a secondary loop (server coolant loop). The CDU is the interface between the two loops. The secondary loop circulates coolant from the CDU through the rack manifold, through hose assemblies and QDC couplings, to the cold plates on each server, and back.
Key Components
CDU — Cooling Distribution Unit
The CDU is the thermal and hydraulic interface between the facility water circuit and the server cooling loop. It contains a heat exchanger, pump, expansion tank, and control system. The CDU maintains the secondary loop at the required supply temperature and flow rate, and provides the pressure head to circulate coolant through the rack manifold and server cold plates. CDU components include heat exchangers, pump assemblies, control valves, and sensors.
CDU ComponentsRack Manifold
The rack manifold distributes coolant from the CDU to each server in the rack. It consists of a supply header and a return header, with branch connections for each server position. Manifolds are typically mounted vertically at the rear of the rack. Each branch connection has a shut-off valve or QDC to allow individual servers to be connected and disconnected without draining the rack loop.
Manifold ComponentsHose Assemblies
Flexible hose assemblies connect the rack manifold to the server QDC couplings. They must accommodate the physical movement of servers being installed and removed, and must be compatible with the coolant chemistry. Hose assemblies are specified by inner diameter, working pressure, temperature range, end fittings, and coolant compatibility.
Hose AssembliesQDC — Quick Disconnect Couplings
QDC couplings are the connection point between the rack manifold hose and the server cold plate circuit. They allow servers to be connected and disconnected from the cooling loop without tools and without significant coolant spillage. Dry-break QDCs seal both halves when disconnected, preventing coolant loss. QDC selection involves connection size, flow rate, pressure drop, coolant compatibility, and mating cycle requirements.
QDC ComponentsCold Plate
The cold plate is the thermal interface between the coolant and the chip package. It is mounted directly on the GPU or CPU, with a thermal interface material (TIM) between the cold plate base and the chip lid. Coolant flows through internal channels in the cold plate, absorbing heat from the chip surface. Cold plate design — channel geometry, material, and flow path — determines thermal resistance and pressure drop.
Cold PlatesMonitoring and Leak Detection
A direct-to-chip cooling system requires monitoring of coolant temperature (supply and return), flow rate, and pressure at key points in the loop. Temperature differential between supply and return indicates the heat load being removed. Flow rate monitoring detects blockages or pump failures. Pressure monitoring detects leaks or restrictions.
Leak detection is critical in data center liquid cooling installations. Coolant leaks near IT equipment can cause electrical damage and data loss. Leak detection sensors are placed at the base of the rack, under the CDU, and at other low points in the coolant circuit. Point sensors detect pooled liquid; rope sensors detect leaks along a cable run. Leak detection systems should be integrated with the data center management system to trigger alarms and automatic shut-off.
Design Considerations
- Coolant supply temperature: confirm the facility water supply temperature is compatible with the server cold plate and CDU operating range.
- Flow rate and pressure drop: calculate the total flow rate required for the rack heat load, and verify the CDU pump can overcome the system pressure drop.
- Coolant chemistry: specify the coolant type (deionised water, glycol mixture, or proprietary fluid) and inhibitor package for the materials in the loop.
- QDC selection: match QDC connection size and flow rating to the server manifold and hose assembly.
- Cold plate interface: confirm cold plate mounting pattern and TIM specification match the server processor package.
- Leak detection: design the leak detection system before installation; retrofit is difficult in dense rack environments.
- Serviceability: ensure hose lengths allow servers to be pulled out for maintenance without disconnecting the cooling loop.
Procurement Considerations
Direct-to-chip cooling systems for AI data centers involve multiple component types from different manufacturers. Procurement teams and system integrators should consider:
- Component compatibility: QDC couplings, hose fittings, and manifold connections must be dimensionally and chemically compatible.
- Lead times: CDU components, custom cold plates, and large-diameter manifolds may have extended lead times — plan procurement early.
- BOM completeness: a complete BOM for a rack cooling loop includes CDU components, manifold assemblies, hose assemblies, QDC couplings, cold plates, sensors, and leak detection.
- OEM and custom requirements: AI server manufacturers may specify proprietary cold plate interfaces or QDC types — confirm before sourcing.
- Volume and project scale: large data center deployments benefit from consolidated sourcing and volume pricing across component categories.