Direct-to-chip liquid cooling

Direct-to-chip liquid cooling, also called cold plate cooling or direct liquid cooling (DLC), removes heat from processors and accelerators by pumping liquid through a metal plate mounted directly on the chip. Water carries heat far more effectively than air, so a cold plate can handle chip power levels that would require impractical airflow and heat sink sizes.

It has moved from high-performance computing niches into mainstream deployment because modern GPUs and CPUs dissipate several hundred watts to over a kilowatt each. For owners and engineers, the shift means the data center now needs a liquid distribution system that reaches every rack, in addition to air cooling for the remaining components.

How cold plate cooling works

A cold plate is a copper or aluminium block with internal microchannels or fins. It sits on the processor lid with a thermal interface material between them. Coolant enters the plate, absorbs heat and leaves warmer. Within the server, tubing connects multiple cold plates in series or parallel, terminating at quick-disconnect couplings at the back of the chassis.

From the server, coolant flows to rack manifolds, then to a coolant distribution unit (CDU). The CDU contains a heat exchanger, pumps, filtration and controls. It transfers heat from the server loop to the facility water system.

ASHRAE terminology describes two loops:

• Technology cooling system (TCS): the secondary loop between the CDU and the IT equipment. It is clean, filtered, low volume and often uses treated water or a propylene glycol mixture.
• Facility water system (FWS): the primary loop between the CDU and the heat rejection plant, such as chillers, dry coolers or cooling towers.

Partial heat capture

Cold plates usually cover processors, accelerators and sometimes memory and voltage regulators. Power supplies, drives, network cards and other components remain air cooled. Depending on server design, cold plates typically capture somewhere around 70 to 90 percent of server heat. The rest must still be removed by room air handling, rear-door heat exchangers or in-row coolers. Designing for the residual air load is a common oversight.

ASHRAE liquid cooling classes

ASHRAE TC 9.9 defines facility water classes by maximum supply temperature to the IT equipment or CDU. The current naming uses the temperature in the class name.

Equipment vendors state which classes their hardware supports, and some newer accelerators specify narrower supply windows than the class maximum. Always design to the vendor's requirements, then use the class to plan heat rejection.

In Canada, supply temperatures of 30 °C (86 °F) or more allow dry coolers to reject heat without compressors for nearly all hours of the year in most regions. That makes direct-to-chip cooling attractive for energy and water savings, and it opens the door to heat reuse.

Coolants and materials

The TCS loop typically uses one of the following:

• Treated deionized or demineralized water with corrosion inhibitors and biocide
• Propylene glycol solutions, commonly around 25 percent (often called PG25), which provide some freeze and biological protection
• Dielectric fluids in some specialised systems

Wetted materials matter. Copper, brass, stainless steel and specific polymers are common. Mixing aluminium with copper in the same loop invites galvanic corrosion unless the coolant chemistry is designed for it. Filtration down to fine levels (vendors frequently specify 25 to 50 microns or finer) protects the microchannels in cold plates, which clog easily.

Fluid quality is not a one-time task. It needs a monitoring program covering pH, conductivity, inhibitor concentration, biological growth and particulates, much like any closed hydronic loop, but with tighter limits.

Facility design considerations

Piping and distribution

Facility water must reach every CDU, and with row-based CDUs that can mean piping in or above every row. Overhead distribution keeps water out of underfloor cable spaces but requires drip trays and careful coordination with cable trays and busway. Underfloor distribution is common in raised-floor rooms. Either way, provide isolation valves, drains, air vents and flushing connections.

Before the IT equipment is connected, the new piping must be flushed and cleaned thoroughly. Construction debris is the single most common cause of early cold plate problems.

Temperatures and delta T

Server cold plates are generally designed for a relatively high temperature rise, often around 10 K, though vendors vary. A higher delta T reduces flow and pump energy. The CDU approach, the difference between facility supply and TCS supply, typically falls in the range of 2 to 5 K. Facility return temperatures can therefore be high enough to make heat recovery useful.

Leak detection and containment

• Use dripless quick-disconnects rated for the operating pressure.
• Install leak detection cable or point sensors in CDUs, under manifolds and along piping.
• Keep TCS operating pressure modest, and in some designs use negative-pressure loops that draw air in rather than leak liquid out.
• Integrate leak alarms into the BAS or DCIM with clear response procedures.

Redundancy

CDU pumps and controls are single points of failure for the racks they serve. Typical arrangements include redundant pumps within each CDU, N+1 CDUs on a shared secondary header, and dual facility water feeds. Cold plates have little thermal mass, so chips can reach throttling temperatures within seconds of losing flow. Continuous cooling on UPS power for CDU pumps is normally required.

Planning steps for owners

1. Get the IT vendor's liquid cooling specification: supply temperature range, flow per kW, pressure drop, fluid type, filtration and connection details.
2. Decide on CDU arrangement: in-rack, in-row or centralised.
3. Determine the residual air-cooled load and how it will be removed.
4. Select heat rejection and plant temperatures based on the class needed.
5. Design piping, isolation, filtration, leak detection and controls.
6. Plan flushing, fluid fill and water quality testing.
7. Commission with heat load simulators before production hardware arrives.

Frequently asked questions

Can direct-to-chip cooling be added to an existing air-cooled data center?

Yes, and it is a common path. The work involves bringing facility water to the rows, adding CDUs and leak detection, and confirming the plant can supply the needed temperature. Our guide to retrofitting legacy data centers covers the steps.

Is direct-to-chip better than immersion cooling?

Neither is universally better. Direct-to-chip fits standard rack form factors and existing operating practices, while immersion captures almost all heat but changes how hardware is handled. Many sites choose cold plates because servers remain serviceable in conventional racks.

What happens if the facility water fails?

The CDU loses its heat sink and the TCS warms quickly. Designs need redundant facility water paths, and controls that alert operators and allow IT load to throttle or shut down safely if cooling cannot be restored.

Key takeaways

• Cold plates move heat from chips directly into liquid, supporting chip and rack powers air cannot handle.
• Systems use a clean secondary loop (TCS) served by CDUs, isolated from the facility water system.
• ASHRAE W classes link supply temperature to heat rejection; warmer water means more free cooling in Canada.
• A significant share of heat remains air cooled and must still be designed for.
• Cleanliness, fluid chemistry, leak detection and pump redundancy decide long-term reliability.