Select a coolant distribution unit against the identified rack's thermal and hydraulic requirements and the facility's available heat-rejection conditions. Confirm coolant compatibility, temperature, flow, pressure, controls and failure behavior together. A nominal kilowatt rating is insufficient unless the supplier shows that the CDU delivers the required duty at the project's actual operating conditions.
A GPU purchase can create a separate cooling procurement task. The rack vendor may define the required coolant conditions while another supplier provides the CDU and the facility team provides the primary loop. Those boundaries need to form one workable system.
SecondWatt's liquid-cooling catalogue supports equipment research. This guide turns an identified GPU rack requirement into a sourcing brief and acceptance worksheet, including the information needed when evaluating used or surplus cooling equipment.
Key Takeaways
- Obtain the exact server or rack cooling requirement before selecting the CDU.
- Compare capacity at the stated fluid, temperature and flow conditions.
- Check the pump duty and pressure limits as carefully as heat-transfer capacity.
- Identify the facility heat-rejection path and any residual room-cooling load.
- Define controls, redundancy, condition evidence and commissioning responsibilities in the offer.
Start with the rack requirement and heat boundary#
Identify the OEM system, configuration and supported cooling arrangement. Request the installation document revision that applies to the offered hardware. GPU model names alone cannot establish coolant requirements, connector compatibility or the proportion of heat captured by the liquid system.
NVIDIA's GB200 NVL72 description identifies a liquid-cooled rack-scale platform. That establishes the broad architecture, not a universal CDU specification. The actual rack implementation supplies the information needed to connect the cooling system.
Rack-side inputs required for CDU selection#
| Input | Required definition |
|---|---|
| Equipment | OEM system, configuration and document revision |
| Liquid heat load | Heat transferred to the coolant at the design operating case |
| Residual heat | Heat remaining for air-side or other cooling |
| Coolant | Approved fluid, concentration and material conditions |
| Temperature | Required supply/return conditions and operating limits |
| Flow | Required flow range and distribution assumptions |
| Pressure | Available/required differential pressure and component limits |
| Connections | Manifolds, couplings and interface requirements |
Do not substitute rack electrical provisioning for the liquid heat load. Some electrical consumption may be rejected outside the liquid circuit, and design margins can describe a different boundary. Ask the rack supplier to identify the load used for cooling selection and how it relates to the intended workload.
Use the GB200 NVL72 dossier or GB300 NVL72 dossier to organize the platform enquiry. The corresponding OEM cooling documents should then govern the component-level selection.
Choose the heat-rejection architecture#
A CDU circulates and controls coolant within a defined system. Its heat exchanger transfers that heat to another medium or circuit. The receiving facility must be able to carry the heat away under the required operating conditions.
CoolIT's CDU portfolio distinguishes liquid-to-liquid and liquid-to-air approaches. The first transfers heat to a facility liquid circuit; the second rejects heat to air. A liquid-to-air arrangement still needs an adequate air-side heat-removal path in the receiving space.
Architecture comparison for a sourcing brief#
| Arrangement | Facility interface to establish | Procurement implication |
|---|---|---|
| Liquid-to-liquid CDU | Primary-loop temperatures, flow and pressure | Confirm the facility can accept the heat |
| Liquid-to-air CDU | Air conditions and room heat-removal capacity | Include the resulting air-side load |
| Dedicated rack CDU | Rack interface and local service arrangement | Define the single-rack operating boundary |
| Shared CDU system | Distribution network and connected-rack duty | Review capacity and failures across the group |
Vertiv's CoolChip CDU 121 page describes an in-rack liquid-to-liquid implementation for single-rack applications. That example illustrates a specific equipment arrangement; its model name or advertised capacity should not be treated as a guarantee at every project's water conditions.
Keep the CDU, primary-loop pumps, heat-rejection plant and rack distribution boundaries visible. A quotation can include one of those systems while excluding the others. Give each exclusion an owner and a corresponding design or procurement task before comparing complete deployment costs.
Calculate an initial flow requirement without mistaking it for selection#
The heat balance provides a useful first check. Heat transfer equals mass flow multiplied by specific heat capacity and the coolant temperature rise. Use consistent units and the actual fluid properties at the relevant conditions when developing the engineering calculation.
The following example is illustrative. Assume a liquid heat load of 240 kW, a water-like fluid with specific heat capacity of 4.18 kJ/kg·K and density of 1 kg/L, and a permitted temperature rise of 10 K. These are calculation assumptions, not an approved coolant specification or a rack vendor's operating limits.
Illustrative flow calculation at two assumed temperature rises#
| Assumed liquid heat load | Assumed temperature rise | Calculated mass flow | Approximate volume flow at assumed density |
|---|---|---|---|
| 240 kW | 10 K | 5.74 kg/s | 344.5 L/min |
| 240 kW | 15 K | 3.83 kg/s | 229.7 L/min |
For the first row, 240 divided by the product of 4.18 and 10 gives approximately 5.74 kg/s. Multiplying by 60 and dividing by the assumed density gives approximately 344.5 L/min. The second row changes only the assumed temperature rise.
The lower flow in the second row does not establish that the rack can accept that larger temperature rise. It also does not address pressure drop, temperature distribution, minimum flow, pump capacity or heat-exchanger performance. Those constraints can determine whether the arrangement is usable.
Replace the assumptions with the approved coolant properties and actual equipment limits. Glycol mixtures and other fluids require their own data. The rack and CDU suppliers should confirm the resulting operating point rather than accepting a water-based estimate as the final selection.
Match heat-exchanger duty and pump performance#
Request a selection sheet that states the primary and secondary supply/return temperatures, flow rates, fluid properties and heat duty. The supplier should show the CDU's performance at those conditions, including the relevant limits and any capacity reduction outside the selected point.
QCT's GB200 cooling paper illustrates a vendor-specific implementation. Treat its operating information as belonging to that design; it is not a source of interchangeable limits for every GPU rack or CDU.
CDU selection sheet to request#
| Selection field | Evidence required |
|---|---|
| Thermal duty | Delivered capacity at the stated design conditions |
| Primary circuit | Fluid, temperatures, flow and allowable pressure conditions |
| Secondary circuit | Fluid, temperatures, flow and available pump duty |
| Pressure losses | CDU and external-network boundaries |
| Operating range | Limits at low load and changing facility conditions |
| Electrical input | Pumps, controls and other supplied loads |
| Failure case | Remaining capability under the specified component outage |
Review the complete secondary flow path. Hoses, manifolds, connectors, filters and cold plates contribute to the required pump duty. A CDU that can transfer the required heat in one test arrangement may not provide sufficient flow through the receiving project's actual circuit.
Check the worst credible facility conditions, not only a convenient nominal point. Warmer primary supply, changing rack demand or an unavailable component can alter the operating result. Ask the supplier to state which combinations are covered and which require a different unit or operating restriction.
Compare the initial installation with the planned expansion. A CDU selected for a future group of racks still needs to operate within its supported range when only the first racks are connected. Ask how flow distribution and temperature control change as branches are added, and identify any commissioning work required at each stage.
Keep the capacity margin explicit. State whether it covers load uncertainty, future equipment, a component outage or another defined condition. Applying several unexplained margins at different boundaries can produce an oversized purchase without establishing the required failure-case performance. Removing every margin can conceal uncertainty just as easily. The selected duty should explain what each allowance is intended to cover.
The final supplier response should therefore show a set of relevant operating cases, not only a single maximum-capacity claim. Use that information to compare units on the service they deliver through the project's expected development and operating conditions.
HPE's GB300 NVL72 QuickSpecs describe a secondary circuit transferring heat through a CDU to a facility primary circuit. That separation is useful for assigning responsibility: each side needs specified conditions, and neither supplier should assume the other side supplies an undefined service.
Verify coolant, materials and physical interfaces#
Obtain written confirmation that the fluid and wetted materials are compatible across the entire connected system. Include the cold plates, CDU, hoses, seals, connectors and manifolds. A fluid accepted by one component supplier may not establish approval for every other component.
Request the required cleanliness and fluid-quality criteria from the applicable equipment documentation. Assign responsibility for filling, sampling, filtration and ongoing monitoring to qualified parties. Do not introduce a generic acceptance concentration or contamination threshold that has not been approved for the actual system.
Fluid and interface evidence#
| Topic | What the buyer should obtain |
|---|---|
| Fluid identity | Approved product or specification and concentration |
| Materials | Compatibility confirmation across the connected circuit |
| Cleanliness | Applicable criteria and commissioning evidence |
| Connections | Correct mating interfaces, ratings and supplied quantities |
| Service access | Space for inspection, maintenance and component replacement |
| Leak management | Detection, alarms and agreed system response |
| Documentation | Final diagrams and identified responsibility boundaries |
Supermicro's accelerator systems demonstrate that server implementations vary in physical and cooling arrangements. Confirm the actual system design rather than ordering a CDU against the accelerator generation alone.
For a used CDU, request service records, previous fluid information, known leakage or contamination events, storage history and completed refurbishment work. Have the relevant supplier assess what cleaning, inspection or component replacement is needed before connecting it to the receiving equipment.
Unused equipment also needs a condition review. Storage duration, preservation, missing connectors or obsolete control components can affect installation readiness. The accepted condition should be supported by evidence and a defined commissioning scope, not only the seller's condition label.
Define redundancy, controls and commissioning#
State the required cooling service during maintenance and component failure. Identify which pumps, controllers, power supplies, CDUs or facility systems are assumed unavailable in each case. A redundant pump inside one unit does not establish redundancy for the entire heat-rejection path.
Cooling operating and failure cases#
| Case | Question to resolve |
|---|---|
| Normal operation | Are all connected racks within their required conditions? |
| Low load | Does control remain stable within the approved operating range? |
| Pump or CDU outage | What flow and heat-removal capacity remain? |
| Facility-loop disturbance | What happens when primary conditions move outside the design point? |
| Electrical disturbance | Which cooling functions continue and for how long? |
| Leak or alarm | What response occurs and who receives the event? |
Connect the cooling controls to the site's monitoring and operating procedures. Agree alarm ownership, trend data, access and escalation. The commissioning plan should demonstrate the required functions under controlled conditions, with any intrusive work performed by qualified personnel using the applicable instructions.
Coordinate electrical continuity with the UPS procurement review. A compute system remaining powered does not establish that its cooling service remains available. The project needs a matched operating and recovery plan for both.
Record the final settings, fluid information, diagrams, test results and outstanding items at handover. The operating team should receive a usable baseline and a maintenance responsibility schedule. A commissioning label without those deliverables can leave the owner unable to distinguish a later change from the accepted condition.
Agree which performance trends the owner will retain after commissioning. Relevant temperatures, flow, pressure and alarms can help compare subsequent operation with the accepted baseline. Identify sensor locations and units so two similarly named measurements are not assumed to describe the same point. Changes to connected racks, coolant or control settings should be recorded with the corresponding engineering review and any required recommissioning work.
Compare offers and submit the cooling brief#
Normalize offers at the same operating point and scope. Include the CDU, pumps, controls, manifolds, hoses, connectors, fluids, installation and acceptance work as applicable. Record the external facility work separately with a named owner and budget status.
Minimum CDU sourcing request#
| Requirement | Information to send |
|---|---|
| Connected equipment | Exact server/rack models and quantities |
| Heat duty | Liquid heat load and residual cooling boundary |
| Coolant | Approved fluid and quality requirements |
| Temperatures | Rack-side requirements and facility-side conditions |
| Hydraulics | Required flow, pressure limits and network information |
| Resilience | Normal, maintenance and defined failure cases |
| Installation | Space, connections, power and control interfaces |
| Commercial scope | Condition, services, destination and target date |
The data-center power procurement roadmap helps align cooling with the wider facility programme. Where electrical capacity changes, coordinate the transformer-sizing review using the complete project loads and boundaries.
SecondWatt acts as an independent intermediary. To request CDU sourcing, send the intended rack configuration, liquid heat load, approved coolant, temperature and flow requirements, facility conditions and target date. Include the available vendor documents so candidates can be compared at a defined operating point.
FAQ: CDU selection for GPU racks#
Can I select a CDU from its kilowatt rating alone?#
No. Capacity depends on the stated fluid, temperatures, flow and other operating conditions. Obtain a selection sheet for the actual project and confirm the pump duty, pressure limits and interfaces. A nominal or advertised capacity without its conditions is insufficient to establish that the CDU can serve the intended rack arrangement.
Is a CDU a substitute for the facility cooling plant?#
Not automatically. It transfers heat within a defined architecture. A liquid-to-liquid unit needs a suitable facility liquid circuit, while a liquid-to-air unit creates an air-side heat-removal requirement. Identify where the heat goes after leaving the CDU and include that system in the project's capacity and responsibility review.
Can I use water-based flow calculations for glycol coolant?#
Use the actual approved fluid properties for the final calculation. The water-like assumptions in this article only illustrate the heat balance. Different fluids or concentrations can change thermal and hydraulic behavior. Have the suppliers confirm the operating point against the rack requirements and the CDU's applicable performance data.
Does a redundant pump make the cooling system redundant?#
It addresses only the functions covered by that pump arrangement. Other components and shared facility systems may remain single dependencies. Define the required failure case and determine the remaining cooling service across the whole path. Compare that result with the rack's operating and recovery requirements before claiming the design meets the intended resilience objective.
Is a used CDU suitable for a new GPU deployment?#
It may be, if compatibility, condition and support can be documented. Review the previous fluid, service and storage history, control configuration and required refurbishment. Confirm the final performance at the new operating point and agree commissioning evidence. A used unit's original application does not by itself establish suitability for a different rack.
What information should accompany the first enquiry?#
Provide the exact rack or server configuration, quantities, liquid heat load, coolant specification, temperature and flow requirements, facility conditions and target date. Add hydraulic information, controls and failure-case requirements where available. Identify unresolved inputs clearly so the response can distinguish a preliminary candidate from a completed equipment selection.