Direct-to-Chip Cooling: Server and Facility Requirements

Direct-to-chip cooling removes heat through cold plates attached to components such as CPUs and GPUs. Coolant flows inside the plates and connected pipework; the electronics are not immersed. A working installation needs a supported server configuration, a controlled coolant circuit, a path to outdoor heat rejection and a separate plan for components that still release heat into room air.

For a buyer, the decisive question is whether the server, rack, cooling equipment and facility share compatible operating limits and clear responsibilities. Selecting a cold plate or a CDU cannot settle those interfaces on its own.

TL;DR

• Begin with the server OEM’s supported configuration and cooling requirements.
• Map the circuit from the cold plates through hoses, rack manifolds and the CDU or other specified heat-transfer arrangement.
• Keep liquid heat and residual-air heat as separate duties. The liquid-cooled fraction depends on the complete equipment configuration.
• Match fluid chemistry, temperature, flow, pressure, cleanliness and connectors across every supplier boundary.
• Plan isolation, leak response, safe replacement and restart before installation.
• Commission at the specified workload and cooling conditions, including approved maintenance and fault cases.

How direct-to-chip cooling works

A cold plate provides a thermal path between the component and circulating coolant. Heat passes through the plate into the fluid, which carries it to a heat exchanger or another specified heat-rejection stage. Supermicro’s technical definition describes this cold-plate, circulation and heat-exchange arrangement.

In a conventional single-phase direct-to-chip circuit, the coolant remains liquid. Hoses and manifolds connect the supported server assemblies to the rack cooling interface. The pump, heat exchanger and controls must maintain the conditions required by the equipment at the intended load.

Cold-plate cooling differs from immersion, where a dielectric liquid touches the electronics. It also differs from a rear-door heat exchanger, which captures heat after air has passed through the rack. These distinctions affect the supported server, maintenance procedure and facility interface. Our liquid-cooling guide compares the broader architectures, while the immersion guide addresses that separate equipment choice.

When should a project consider direct-to-chip?

Consider it when the selected server requires cold-plate cooling, or when procurement is still open and a supported liquid-cooled configuration could meet the workload within the site’s power, space and cooling limits. There is no universal rack-kilowatt threshold at which every server must change architecture.

An existing air-cooled server is not automatically a candidate for a retrofit. Obtain the OEM’s support position for the complete conversion, including cold plates, hoses, firmware, materials and warranty responsibility. A component’s thermal result does not establish support for the assembled server.

For a new deployment, compare supported server and cooling combinations at the same useful workload. Examine infrastructure cost, auxiliary energy, service access, spares and the site heat-rejection options. Direct-to-chip can change the facility design substantially, but its name alone does not establish a whole-system efficiency advantage for every project.

Follow the coolant circuit across supplier boundaries

The Open Compute Project cold-plate workstream treats the technology cooling system as a combination of cold plates, tubing, manifolds, connectors, CDUs and working fluids. A buyer needs that complete path documented, including its interfaces with the facility.

Many liquid-to-liquid installations use a CDU to exchange heat between the equipment-side coolant and the facility-water circuit. Keeping the circuits separate allows their fluid and operating conditions to differ within the exchanger’s design limits. It does not mean either side can operate at arbitrary conditions.

Identify the actual piping arrangement before assigning duties. A liquid-to-air CDU, for example, releases heat into air rather than transferring it into a facility-water circuit. Vertiv’s liquid-to-air CDU announcement describes that retrofit route. The building then needs an air-side path capable of carrying that heat to outdoor rejection.

Draw each heat path and label the measurement points. State where temperature, flow and pressure limits apply, where sensors are installed and who receives each alarm. Our CDU sizing guide covers the thermal and hydraulic selection; the requirements here establish the equipment and facility inputs for that calculation.

Liquid cooling still leaves residual-air heat

Cold plates cool the components included in the supported design. Other equipment may reject heat to air, including parts of the server, power conversion, networking or storage. Ask for the complete rack’s liquid and air heat duties at the intended operating conditions.

Consider an illustrative rack with 100 kW of IT input, assuming that input becomes heat within the site boundary. Choose 80 kW into liquid and 20 kW into room air for this example:

Total heat = 80 kW liquid + 20 kW air = 100 kW.

The 80/20 split is a chosen illustration, not a typical value or a specification for a named server. It shows why a liquid circuit rated for 80 kW would leave a separate 20 kW air-side task. For four identical racks under these assumptions, the duties would be 320 kW liquid and 80 kW air, totaling 400 kW before adding cooling auxiliaries and other facility heat.

If the equipment transfers less heat to liquid, the room-air duty rises. Use the OEM’s worst supported operating case, including any changes associated with workload, firmware or equipment population. Do not infer the split from GPU power alone or reuse a percentage from another rack configuration.

Cooling redundancy must cover the appropriate duty during the agreed maintenance and fault cases. A liquid circuit that remains available does not compensate for losing the room-air system if the remaining components exceed their permitted inlet conditions.

Specify the interfaces in one shared schedule

Require values at identified locations and conditions. “Warm-water compatible” is insufficient without the supported temperature range, flow requirement and heat-exchanger performance at the selected site conditions. Likewise, a maximum CDU capacity measured at one temperature difference may not be available at another.

Resolve connector compatibility at assembly level. Similar appearance or a common connector category cannot establish mating compatibility, permitted pressure or an approved service procedure. Obtain the supplier’s exact part references and supported pairings.

Fluid quality and condensation belong in the design

Use the equipment suppliers’ agreed coolant specification. Record allowable chemistry, treatment, cleanliness, sampling and replenishment practices, together with the wetted materials. A circuit can contain cold plates, flexible hoses, seals, manifolds and heat exchangers from several vendors; their requirements need to be compatible.

Do not treat ordinary building water as an approved server coolant. Flushing, filling and fluid sampling need a documented procedure and acceptance criteria. Establish who supplies the initial charge, who checks it after installation and who investigates a later fluid-quality alarm. Vertiv’s CDU application and planning guide illustrates how an equipment-side circuit has its own fluid, filtration and control requirements.

Check condensation risk wherever a coolant surface can be colder than the surrounding dew point. The design may use temperature control, insulation or other approved measures. Specify sensors, control limits and the behavior during changes in room humidity or cooling supply; a nominal water temperature is insufficient to establish protection under every condition.

The equipment-side circuit and the outdoor heat-rejection method are separate choices. Direct-to-chip cooling does not establish that a facility consumes no water. Our guide to liquid cooling and water use distinguishes closed circuits, dry coolers and evaporative systems. Site climate and the required cooling temperatures also shape the selection, as discussed in our extreme-heat cooling guide.

Design for the intervention the operator will perform

A maintenance plan should show how an engineer isolates a server or branch, disconnects it, manages any released fluid and returns the equipment to service. Reserve space for hose movement, tools, replacement assemblies and any required handling aids. Check the route with adjacent racks installed.

Use the supported disconnect and reconnection procedure. Do not infer that a quick-disconnect label authorizes hot service under every operating condition. The equipment supplier must define the isolation state, permitted intervention and restart checks.

Agree the response to a detected leak, blocked filter, loss of circulation and sensor fault. Which alarms go to the rack controller, BMS or operator? Who can command a pump, valve or workload change? Which faults require isolation, load reduction or shutdown? These answers depend on the selected system and its approved controls.

Stock the specified replacement components and fluid, and train the operators on the actual installation. Record the escalation route across server, rack, CDU and facility suppliers. A leak alarm without an agreed responder and isolation procedure leaves the operational design unfinished.

Commission at the supported operating conditions

Before energizing IT, confirm circuit cleanliness, fluid quality and the agreed pressure/leak-test results using the approved procedures. Avoid assigning a universal test pressure: each installed component and assembly needs an appropriate documented limit.

During the workload test, measure server or rack input, coolant inlet and return conditions, branch flow where required, component temperatures and room-air conditions. Record useful workload and any throttling or errors. The test should show that thermal conditions remain within the equipment’s supported limits at the specified design case.

Demonstrate representative service and fault cases through supplier-approved procedures. Include loss of an available cooling component, relevant alarm paths and restart checks. Identify which cases were physically tested and which were established through other evidence. Carry open items into the handover record with an owner and closure requirement.

ModulEdge’s role in a direct-to-chip facility

ModulEdge’s MDC2000 approach combines direct-to-chip interfaces with a separate residual-air cooling provision. It is relevant when the buyer wants a modular facility developed around dense liquid-cooled AI equipment, with coordinated electrical, thermal, controls and service interfaces.

The chosen OEM rack determines the detailed cooling conditions and equipment-side arrangements. The facility design must follow that configuration rather than transfer a published temperature, rack count or thermal split from another platform.

MDC500 serves a different documented architecture: air-cooled IT with active rear-door heat exchangers. That distinction helps a buyer select the facility route around the supported server configuration. Share the equipment schedule and thermal requirements before comparing the two approaches.

Frequently asked questions

Is direct-to-chip cooling the same as immersion?

No. Direct-to-chip coolant stays inside cold plates and connected pipework. Immersion places supported electronics in a dielectric liquid.

Does direct-to-chip cooling remove all heat through liquid?

The fraction depends on the complete server and rack configuration. Obtain separate liquid and residual-air duties and design a cooling path for each.

Can I retrofit any air-cooled server with cold plates?

Only a supported conversion can establish the equipment configuration and warranty responsibility. Obtain the OEM’s position on the complete assembled system before procuring a retrofit.

Do I always need a facility-water circuit?

The heat-transfer arrangement determines that. A liquid-to-liquid CDU connects to a facility circuit; a liquid-to-air arrangement releases heat into air, which the building must then remove.

Can I use ordinary water in the server circuit?

Use the agreed equipment-side coolant specification, including quality, treatment and material compatibility. Do not assume untreated building water meets it.

What should be tested before handover?

Test the approved thermal operating case, fluid and leak acceptance requirements, controls and alarms, representative maintenance and agreed fault responses. Record the conditions and any cases still awaiting evidence.