Engineering guide 01 / 08
Liquid Cooling Design and Specification for MEP Engineers
A coolant distribution unit (CDU) can have enough cooling capacity and still be a poor fit for the installed racks. Start with the connections, fluid and operating conditions, then turn those decisions into a specification the commissioning team can prove.
On this page
A complete MEP liquid cooling specification connects the facility water system (FWS), CDU and technology cooling system (TCS) serving the racks. Define the permitted fluid, flow, temperature and pressure at each interface, then state who verifies cleanliness, performance and readiness before handover.
- For
- MEP consultants and data-center design engineers
- Scope
- Single-phase, water-based direct-to-chip systems using liquid-to-liquid CDUs. Liquid-to-air, refrigerant, immersion and two-phase systems require different interface and fluid requirements; apply only the relevant principles after checking their equipment documentation.
Key decisions
- Resolve conflicting CDU, rack and coolant requirements before issuing the final specification.
- Check duty at actual fluid properties, available pressure and the design temperature envelope.
- Specify access and evidence for cleaning, sampling, monitoring and maintenance during design.
- 01Facility Water System
Heat rejection, supply conditions and facility-side treatment
- 02CDU heat exchanger
Heat transfer across separate facility and server fluid circuits
- 03Technology Cooling System
Approved coolant, pumps, filtration and distribution
- 04Rack and cold plates
Verified flow, temperature, pressure and material compatibility
- 05Acceptance record
Measured performance, clean-loop evidence and operating baseline
Arrows show the order of design checks, not a shared fluid path. The FWS and TCS exchange heat through the CDU heat exchanger. Their fluids do not normally mix; operating limits and monitoring must be specified separately.
Will the CDU meet duty under your actual conditions?#
A quoted kilowatt capacity applies at stated conditions. Ask what the CDU delivers at your facility supply temperature, selected coolant and required flow. Review partial load and the planned maintenance state too. If the availability design allows one unit or pump to be out of service, check the remaining capacity in that state.
Confirm the heat-exchanger approach: the temperature difference needed to transfer heat between the loops. It affects how warm the rack supply becomes for a given facility supply. Retain the pumping requirements and support-equipment electrical load with the duty selection.
Liquid cooling may leave a room air-cooling load. NVIDIA documents hybrid cooling in its DGX GB200 architecture. Obtain the liquid-captured heat and remaining air load for the selected equipment. Coordinate coolant temperature with room humidity and dew point, including the approved response if condensation conditions develop.
| Parameter | State the requirement | Verify at design review |
|---|---|---|
| Thermal duty | Liquid heat load and residual room heat for each deployment phase | Actual IT configuration, not electrical capacity alone |
| Temperature | Permitted supply and return ranges for each loop | OEM envelope, heat-exchanger approach and room dew point |
| Flow | Required rack and branch flow across the operating range | Fluid properties and branch balancing |
| Pressure | Static, differential, transient and component limits | Pump curves, expansion arrangement and weakest rated part |
| Coolant | Named formulation, permitted concentration and quality requirements | All wetted components and mixed-vendor approval |
| Filtration | Required efficiency, capacity and loaded-filter pressure allowance | Cold-plate requirements and maintainable filter arrangement |
| Availability | Capacity and service options for specified failure cases | Demonstrated remaining duty and isolation sequence |
References: NVIDIA: DGX SuperPOD Architecture Featuring DGX GB200
Can the required flow reach the most demanding branch?#
Two checks answer different questions: how much fluid carries the heat, and how much pressure moves that fluid through the circuit. Use a heat balance to check consistency with the rack supplier's flow requirement. It does not prove local cold-plate performance or replace minimum-flow limits.
Calculate the loss along the index circuit, the path that sets the required pumping pressure. Include piping, fittings, connectors, manifolds, filters, valves and cold plates. Compare the result with CDU pressure available at the required flow, including the specified loaded-filter condition.
Use density, specific heat and viscosity for the actual coolant at operating temperature. Keep static pressure, differential pressure, component ratings and pump suction constraints distinct. Review expansion and air management with suppliers, and retain the valve positions and control mode that the calculation assumes.
How much water carries a 500 kW heat load?
For a 500 kW liquid heat load and a 10 K temperature rise, water with a specific heat of 4.18 kJ/(kg K) needs a mass flow of approximately 11.96 kg/s. At a density of 1,000 kg/m3, that is about 718 L/min. Check the rack's minimum-flow requirement and pressure losses before selecting equipment.
mass flow = heat load / (specific heat x temperature rise) = 500 / (4.18 x 10) = 11.96 kg/sFor an approved glycol mixture, replace both fluid properties with supplier values. Confirm rack flow and the full pressure-loss calculation separately. The FWS and TCS need not have the same temperature rise or flow.
Which coolant is approved for every wetted component?#
Name the fluid formulation and supplier, permitted substitutes, concentration method and makeup fluid. Include batch, storage and test requirements. A label such as deionized water or a glycol percentage does not establish compatibility with every additive, metal, seal or factory-filled component.
List the parts that touch the fluid, including temporary ones: hoses, valve seats, seals, connectors, joining compounds, sensors, fill carts and test equipment. Obtain approval for chemical cleaning or passivation against that list. A treatment appropriate for one material may be unsuitable for another.
Ask filter suppliers for removal efficiency, capacity and clean and loaded pressure loss. A micron label is only part of the specification. Also document bypass and replacement arrangements. Keep chemistry and particle acceptance separate: acceptable pH or conductivity cannot establish that construction debris has been removed.
Can the installed loop actually be cleaned and sampled?#
Imagine finding the right place for a sample after the racks are installed, only to discover there is no accessible port. Avoid that problem in the coordinated layout. Provide isolation, drains, vents and temporary flushing connections where the selected procedures need them, with room for filter service and spill containment.
The Open Compute Project (OCP) modular TCS guidance covers connection methods for sampling, venting and air removal when IT equipment joins the loop. Review the chosen method with rack and CDU suppliers before finalizing the layout.
Match the sampling location to the question. A circulating return sample can describe mixed bulk fluid while missing an isolated branch. A side-stream monitor represents the fluid reaching it. Record its filter relationship, sample-flow verification and behavior during isolation. A normal bulk reading does not prove a cold plate has no deposits.
What decision follows each monitored change?#
For every point, specify its location, units, range, timestamp, validity indication, verification method and history retention. Then name the alert owner and the next check. Chemistry acceptance limits come from the approved coolant and equipment requirements; investigation thresholds need comparable operating conditions.
Retain the commissioned baseline with load, temperature, flow, pressure, filter condition and sample results. A pressure change during a pump-mode change means something different from the same change under a steady configuration. Context makes the trend interpretable.
Keep diagnostic monitoring distinct from protective controls. A coolant-condition signal can support investigation. Any automatic shutdown role requires its own approved and tested design; the CDU and specified protection systems keep their assigned responsibilities.
What must the commissioning team prove before handover?#
Write the release gate into the specification: sensitive equipment connects only after the designated authority accepts clean-loop evidence. Require the flushing drawing, procedure-water approval, branch coverage, cleaning-residue assessment, final-fluid verification and a plan for delayed startup. State what a failed result or rework triggers.
Keep the same requirements matrix through procurement, installation and acceptance. Identify supplier deviations before delivery. Hand over as-built piping and instrumentation diagrams (P&IDs), fluid batches, test records, instrument checks and procedures with named owners.
Include operators in load tests, maintenance demonstrations and agreed failure scenarios. They should see the accepted operating state and practice the response before the system carries production work. The specification starter below gives the team a common record to complete.
Project specification starter
SYSTEM: [TCS identifier and served racks]
AUTHORITY: [approved design basis, OEM documents and revisions]
FLUID: [supplier, formulation, concentration and approval references]
ENVELOPE: [temperature, flow, static pressure and differential-pressure limits]
CLEANLINESS: [approved method, particle assessment and acceptance limits]
MONITORING: [locations, verified measurements, trend retention and alert owners]
HOLD POINT: No connection to sensitive equipment until [designated authority] accepts [required test records].
HANDOVER: Provide as-built P&IDs, baseline trends, laboratory results, maintenance procedures and training records.Common questions
What does MEP mean in liquid cooling projects?
MEP means mechanical, electrical and plumbing. In data-center liquid cooling, MEP engineers coordinate heat rejection, distribution, power, controls, water treatment and equipment interfaces with IT and specialist suppliers.
Are facility water and technology coolant the same?
Not in the liquid-to-liquid CDU architecture covered here. The FWS and TCS are separated by a heat exchanger and may use different fluids, flow rates and quality requirements. Verify the actual project architecture before applying this distinction.
Can an MEP specification prescribe one pH or conductivity limit for every rack?
No. Obtain the approved coolant and equipment requirements for the selected system. Fluid formulation, additives, concentration, temperature and wetted materials affect interpretation. Record the test method and units alongside every limit.
Does a coolant monitor replace laboratory testing?
No. Continuous condition trends can identify changes between samples and provide operational context. Laboratory testing remains complementary for the approved chemistry, contamination and microbial analyses needed by the project.
What should be included in a liquid cooling design handover?
Provide approved interface documents, as-built schematics, fluid and material approvals, cleaning records, pressure and performance tests, instrument verification, laboratory results, baseline trends, maintenance procedures and training records.
Sources and further reading
Reliability Engine
Connect coolant condition to operating decisions
Reliability Engine combines side-stream coolant-condition information with available CDU, loop and GPU context to support investigation. Discuss sampling access, verified measurements and operational data requirements during design; laboratory testing remains part of the fluid verification plan.