Commissioning / Field operations
Commissioning Liquid-Cooled AI Racks: A Field Checklist
Your rack may be installed and powered, but is its liquid-cooling system ready for production? Use a staged acceptance record instead of a single “no visible leaks” check.
Sections in this article
The contractor reports that a high-density liquid-cooled rack is installed. Hoses are connected, the CDU screen is green and the system appears ready. Before approving production, the commissioning team still needs evidence that the loop is clean, fluid-compatible, correctly instrumented and able to recover from specified failures.
Use this sequence as a commissioning planning checklist, not as a standardized pressure-test procedure. Test media, pressure, duration, chemistry and acceptable response times must come from the manufacturer, the approved engineering specifications and competent site safety personnel. Some pressure-test techniques are inappropriate for sensitive IT components.
Define precisely what the commissioning boundary includes
Start with a marked-up drawing of the facility water system (FWS), heat exchanger/CDU and technology cooling system (TCS), through to manifolds, hoses and cold plates. Record who owns each interface. A pass at the CDU skid alone is not a pass for an entire rack row.
Approved temperatures, pressures, fluid types, material list, sensor ranges, redundancy.
Vendor functional tests, mechanical integrity, alarms, design-point performance.
Correct connections, approved flushing, leak detection, controls, hydraulic balance.
Approved fault scenarios, operator response, alarms and protective actions.
ASHRAE describes the FWS/TCS distinction and common CDU components in its data center engineering guidance. Keep those interfaces explicit in your test sheets, even when a single vendor supplies most of the system.
Before delivery: documents and factory acceptance testing
Request the latest piping and instrumentation diagram, approved fluid specification, list of wetted materials, pump performance curves, heat-exchanger rating at design temperatures, filter requirements, sensor calibration evidence and alarm/control matrix. Document which interfaces still depend on the server integrator, the facility engineer or the BMS contractor.
For factory acceptance testing (FAT), confirm which functions were physically tested and which were simulated. A 500 kW thermal label may come from a specified factory test condition; it does not establish site performance with warmer supply water or a different coolant. Have the vendor provide measured results and test conditions, not a generic certificate.
- CDU tag
- CDU-DH3-01 (illustrative)
- Approved fluids
- Exact product and concentration
- Recorded capacity
- At stated FWS/TCS temperatures
- Pump test
- Flow versus available differential pressure
- Open items
- Owner, action and due date
Site acceptance: check what was actually installed
First confirm mechanical completion: correct flow directions, hose routing, coupling engagement, pipe supports, isolation valves, strainers, drains, vents, instrumentation orientation and accessible service clearances. Photos and tagged punch-list items make later troubleshooting much easier.
Fluid compatibility must be confirmed as a system: the server cold plates, seals, pumps, fittings and water-treatment package. A generic rule to ban all brass or mandate one conductivity value is not defensible across all designs. Follow the approved bill of materials, supplier specifications and warranty conditions.
Before putting sensitive IT hardware into the path, approve a contamination-control plan. Where appropriate, use bypass loops and temporary filtration for flushing. Define cleanliness criteria and sample locations based on component sensitivity; neither “50 microns for four hours” nor a visually clear sight glass is a universal acceptance criterion.
| Site test | Evidence to retain | Who accepts it |
|---|---|---|
| Loop identity and isolation | Marked-up drawing, valve positions, tag photographs | Mechanical lead |
| Fluid compatibility and cleanliness | Approved chemistry, material register, sample report | Vendor / water specialist |
| Leak / pressure integrity | Approved method, pressure-temperature log, exceptions | Qualified test engineer |
| Flow at design duty | Meter readings, pump speed, loop differential pressure | Commissioning agent |
| Alarms and leak detection | Alarm event IDs, annunciation path, reset sequence | Controls / facilities |
| Rack thermal duty | Temperatures, IT heat load and stable operation | IT + facilities |
Pressure testing needs particular caution. Compressed gas stores significant energy, while excessive hydrostatic pressure can damage delicate cold-plate channels or quick disconnects. Use only an approved test method with appropriate isolation, personnel protection and supplier agreement. Do not copy a fixed 1.25× or 1.5× pressure multiplier into a site runbook without validation.
Integrated systems testing: confirm behavior during faults
Integrated systems testing (IST) should assess how the CDU, cooling plant, rack and operational alerting behave together. Begin with simulation or safe test facilities, then conduct only risk-assessed physical tests under an approved plan. Do not deliberately interrupt cooling to live production GPUs to satisfy a checklist.
Observe the approved sequence, flow response, alarms and whether the remaining system meets the required design duty.
Verify local control continues as designed. A green central dashboard is not proof that field alarms survive a network failure.
Confirm alert routing, safe isolation sequence and an operator-readable location identifier.
Exercise the agreed escalation path and safe workload controls in a simulated or otherwise approved scenario.
Record what happened, not only PASS or FAIL. For example, a backup pump may eventually recover the flow but take longer than the server-approved thermal buffer permits. Actual measured transition time should be compared with a site-specific acceptance criterion.
A short, useful commissioning sign-off
Make the final handover pack easy for operators to use. It should contain the current as-built diagram, setpoints and limits, measured results, alarm response matrix, validated operating modes, emergency contacts, water maintenance plan, spare-parts list and documented open items. Keep superseded drawings out of the active control room copy.
System: CDU-DH3-01 and associated rack secondary loop.
Approved duty: per design specification and measured conditions.
Exceptions: none / attached punch list.
Production authorization: requires signatures from facilities, IT owner and commissioning lead.
Review triggers: rack additions, coolant changes, pump changes or a related incident.
Before signing: ask whether the receiving team can explain the alarm, find the current limits, isolate the affected branch safely and locate the evidence that the system passed its tests. If not, the commissioning package is incomplete even if every checkbox is marked.
Sources and further reading
Primary frameworks and reference materials. Engineering decisions require site-specific validation; legal classifications require assessment of the actual use case.