Liquid Cooling_
Condensation & Dew-Point Control in Liquid-Cooled GPU Halls
Details the engineering steps, sensor placements, and control logic required to keep liquid-coolant supply temperatures above hall dew point in NVL72-class deployments, showing why warm-water loops reduce condensation risk without secondary chillers.
Key facts
- The facility water loop must remain above the measured dew point of the surrounding air to prevent condensation on cold plates and manifolds.
- Warm-water cooling targets supply temperatures that inherently stay above dew point in conditioned halls.
- CDU secondary loops use heat exchangers sized per the OEM approach temperature to keep facility water above hall dew point under worst-case humidity.
- Continuous dew-point calculation requires both dry-bulb and relative-humidity sensors placed at rack inlet height and at CDU inlet.
- Factory-terminated MPO trunks carry only the InfiniBand or Ethernet scale-out fabric; they are never part of the liquid-coolant circuit and do not affect condensation risk inside the rack.
- OEM rack manifolds include drain valves and drip pans sized to capture condensate from a single cold-plate leak before it reaches power distribution units.
- GPU-to-GPU NVLink connections run over the copper spine inside the rack and have no bearing on facility-water dew-point control.
Dew-Point Measurement Placement
Install combination temperature and humidity sensors at multiple elevations on the cold aisle face of each row. Average the readings to compute instantaneous dew point for the CDU control loop.
Mount an additional sensor array inside the contained hot aisle at the same elevations to detect recirculation that could raise local dew point near the rack exhaust. Feed both values to the building management system so the CDU can select the higher of the two dew points as its safety floor.
Calibrate all sensors against a reference instrument at commissioning and on a recurring schedule; drift beyond the stated tolerance triggers replacement before the next maintenance window.
Supply Water Temperature Control Logic
Program the CDU variable-speed pumps and mixing valves to hold facility-water supply temperature at the greater of the OEM-specified minimum or the calculated dew point plus the required margin. This margin accounts for sensor uncertainty and transient humidity spikes during door openings.
When outdoor conditions allow, enable free-cooling towers or dry coolers to raise supply temperature into the warm-water band rather than sub-cooling. The higher temperature reduces chiller lift and keeps the entire secondary loop safely above dew point without additional reheat.
Log supply temperature, dew point, and valve position at regular intervals; any excursion below the safety floor must generate an immediate alarm and automatically open the mixing valve to the hottest available source.
Warm-Water Loop Design Choices
Size the primary-to-secondary heat exchanger for the OEM-specified approach at design load so that facility water never drops below the dew-point floor even when the primary loop is at its coldest seasonal condition. This approach eliminates the need for a separate trim heater on the secondary side.
Route the facility-water supply header above the rack manifolds with a continuous upward slope toward the CDU; any air or vapor pockets rise to automatic vents rather than collecting at the lowest point where condensation could form.
Specify gaskets rated for continuous warm-water operation; lower-rated materials harden and leak when the loop is kept warm, creating a new condensation source on the exterior of the fitting.
Common Failure Modes and Detection
The most frequent field failure is a humidity sensor drift that under-reports dew point, allowing the CDU to supply water below the actual condensation threshold for weeks until visible moisture appears on cold plates. Cross-check every sensor against a portable reference unit during each scheduled PM.
Another common mode occurs when operators lower the CDU set point to chase GPU delta-T during a firmware update that increases power draw; the water temperature falls below dew point before the control loop can react. Add a hard interlock that refuses any set-point change below the current dew-point floor.
Leaking quick-disconnect couplings on the rack side allow small amounts of water to evaporate, locally raising humidity and dew point inside the containment; the resulting condensate then forms on adjacent uninsulated piping. Require a pressure-decay test on every coupling after each rack service event.
Commissioning Sequence
First verify that all rack cold plates and manifolds have been pressure-tested and dried to the required dew point before any facility water is introduced. Only then enable CDU flow with the supply temperature held above the highest measured hall dew point.
Run the hall at maximum design humidity for several hours while monitoring every cold-plate surface with an infrared camera; any spot below supply temperature indicates an insulation gap or sensor error that must be corrected before full load is applied.
Document the final control set points and interlock thresholds in the as-built record so that future capacity additions cannot inadvertently lower water temperature without a corresponding dew-point review.
Ongoing Maintenance Requirements
Leviathan Systems crews replace humidity sensors on a fixed cycle regardless of calibration status because drift accumulates faster in the high-airflow environment of GPU halls. Each replacement is logged against the rack row and CDU it serves.
Quarterly visual inspection of all exposed facility-water piping includes checking insulation integrity at every support hanger; compressed insulation at contact points is the leading cause of localized condensation after the first year of operation.
Annual full-system dew-point mapping repeats the multi-elevation sensor survey while the hall is at both minimum and maximum expected occupancy; any location that now exceeds the original design dew point triggers a containment or airflow adjustment before the next cooling season.
Standards referenced: ASHRAE TC 9.9 thermal guidelines for data processing environments · OCP liquid cooling cold-plate and manifold interface specification
Frequently asked_
What is the minimum supply-water temperature we can safely run in a conditioned hall?
Calculate dew point from the local dry-bulb and RH sensors, then add the safety margin required by the CDU control logic. Supply water should not drop below that value.
Does warm-water cooling change the required flow rate through the GPU cold plates?
Flow rate is set by the GPU OEM thermal specification and remains independent of supply temperature. The higher inlet temperature simply reduces the available delta-T, so the CDU must increase flow or accept a higher GPU outlet temperature within the allowed envelope.
How often must we re-validate dew-point sensors after initial commissioning?
Re-validate against a reference instrument on the schedule defined by the site procedures and replace on the established cycle. Any sensor showing drift beyond tolerance is replaced immediately rather than recalibrated in place.
Can we use the same CDU for both warm-water GPU loops and colder storage or network racks?
Separate secondary loops are required. Mixing a warm GPU loop with a colder loop in one CDU defeats the dew-point protection for the warmer circuit and creates condensation risk on the colder manifolds.
What happens if a quick-disconnect leaks during a rack swap?
The rack-level drip pan captures the initial spill, but evaporating water raises local humidity. The BMS dew-point sensors inside containment will alarm, triggering an automatic increase in CDU supply temperature until the coupling is replaced and humidity returns to baseline.