Commissioning liquid-cooling CDUs without breaking your reliability envelope

Direct-to-chip liquid cooling changes the failure surface of a data hall in ways that traditional CFD models understate. The coolant distribution unit (CDU) becomes a single point of failure for a rack row, and secondary-loop dielectric quality drifts on a timescale that BMS alarms typically miss.

On our last three retrofits in Tseung Kwan O and Tsuen Wan we required the following before energising GPU load.

One: leak detection at every quick-disconnect, wired into the same alarm path as fire and EPO — not a silent BMS point. Two: dual-redundant CDU control with automatic isolation of a failed primary. Three: secondary-loop conductivity trending with hard thresholds, not spot checks at handover. Four: differential-pressure alarms tuned to the rack manifold, because CDU discharge pressure alone hides manifold restriction. Five: coolant-temperature ride-through testing under a simulated chiller trip. Six: a rehearsed manual bypass procedure with two named operators per shift.

None of this is exotic. It is, however, rarely present in the commissioning packs handed over by general contractors. Most packs prove the loop fills, flows and cools under steady state. They do not prove the hall survives the failure modes that matter under OR-2 and tenant SLAs.

We also insist on writing CDU and manifold assets into the technical hierarchy before first GPU watt is drawn. If the asset is not in the hierarchy, it will not appear in criticality or FMEA — and it will not appear in the evidence pack when someone asks.

If you are mid-retrofit, stop and score the six points above. Closing gaps before energisation is cheaper than explaining them after the first excursion.