Liquid Cooling
Air cooling stops being economical somewhere between 15 and 20 kW per rack. The accelerator racks now going
into AI training halls run at 40, 60 and occasionally more than 100 kW, and the only practical way to move that heat is in
liquid, close to the chip.
Liquid cooling is therefore no longer a niche for supercomputers. It has become a construction discipline in its
own right: a wetted circuit designed, welded, tested and commissioned to the same standard as any pressure system, with
coolant chemistry, filtration, leak detection and service access all planned before the first rack is delivered.
HDmann works in that discipline. We build the secondary loop - CDU, rack manifolds, hose assemblies, valves and
fittings - and we interface it with the facility water system and with the rack integrator, so responsibility for the
whole circuit sits in one place.
Material & finish selection
| 316L stainless steel, orbital welded and pickled - wetted headers and manifolds |
| Peroxide cured EPDM (EHW194) - hose and seals in contact with coolant |
| PTFE seats for isolation valves, EPDM for balancing valves |
| PG25 or deionized water, filtration 25 to 50 micron |
Key engineering questions
| Design heat load per rack and per cold plate, and the allowed coolant temperature rise |
| Supply temperature against room dew point - condensation risk on uninsulated surfaces |
| Coolant chemistry: PG percentage, conductivity, inhibitor, make-up water quality |
| Leak detection strategy and the response it triggers at rack and at room level |
| Service access: can a single server be swapped with the row still running |
| Interface definition with the rack integrator (UQD or UQDB, float allowance) |