Liquid Cooling_
Two-Phase vs Single-Phase Direct Liquid Cooling for GPUs
Details field procedures for deploying and maintaining single-phase versus two-phase direct liquid cooling on H100 through GB300-class GPU racks, including manifold routing, leak testing, and failure isolation steps used by crews performing rack assembly and commissioning.
Key facts
- Single-phase DLC circulates water-glycol mixtures through cold plates and returns heated fluid to a coolant distribution unit for rejection to facility chilled water or dry coolers.
- Two-phase DLC relies on refrigerant evaporation at the cold plate and condensation in a remote heat exchanger, requiring sealed loops and level controls for the working fluid.
- Manifold connections in both systems use quick-disconnect couplings rated for the selected coolant; single-phase loops tolerate minor air entrainment while two-phase loops do not.
- GPU cold plates mount directly to the processor lid with a specified thermal interface material; torque sequences follow the GPU OEM cold-plate kit instructions.
- Leak testing for both systems begins with a low-pressure nitrogen hold followed by helium sniffing or pressure decay measurement before introducing coolant.
- Two-phase systems require the condenser to be mounted above the evaporators to maintain proper liquid return; single-phase loops allow more flexible CDU placement.
- Field crews perform visual inspection of every quick-disconnect and O-ring surface with a borescope before final rack insertion.
Single-Phase Cold Plate Installation Sequence
Mount the cold plates to each GPU using the OEM-supplied torque sequence and thermal interface material. Route the supply and return hoses from the rack-level manifold to the cold plates while maintaining minimum bend radii specified by the hose manufacturer.
Connect the hoses to the manifold quick-disconnects after confirming each coupling is clean and the O-ring is seated. Purge air from the loop at the highest point using the installed bleed valves before starting the CDU pump.
Verify flow through each branch with the CDU flow meters; imbalance indicates a pinched hose or incomplete air removal.
Two-Phase System Integration Steps
Position the condenser above the rack evaporators so gravity returns liquid refrigerant. Install the sealed refrigerant lines with the correct slope and support spacing to prevent liquid traps.
Charge the loop only after a triple-evacuation and helium leak test confirms integrity below the OEM allowable rate. Add the precise refrigerant mass stated on the system nameplate.
Confirm that the liquid-level sight glass or sensor shows proper fill before enabling the compressor or pump that drives the condenser loop.
Manifold and CDU Connections
Both architectures terminate at rack manifolds that tie into facility supply and return headers. Single-phase manifolds use larger-diameter pipes because sensible heat capacity requires higher mass flow.
Two-phase manifolds carry lower volumetric flow but must remain perfectly sealed; any non-condensable gas accumulation raises system pressure and reduces cooling capacity.
Leviathan Systems crews label every quick-disconnect pair with loop identifier and direction before final torque to prevent cross-connection during later service.
Routine Maintenance Differences
Single-phase loops require periodic coolant sampling for pH, conductivity, and particulate count; filters are changed when differential pressure across the CDU strainer exceeds the OEM threshold.
Two-phase loops need far less fluid maintenance but demand annual inspection of the condenser coils for airflow restriction and verification that the expansion device maintains the design superheat or subcooling.
Both systems require quarterly visual checks of all flexible hoses for abrasion where they pass through cable trays or rack doors.
Common Failure Modes in Field Installations
The most frequent issue is improper quick-disconnect engagement that leaves a partial seal; this produces slow coolant loss in single-phase loops and immediate capacity loss plus contamination in two-phase loops. Crews catch it by performing a final pressure-decay test after every rack insertion.
Air pockets in single-phase loops cause localized hot spots on GPUs; two-phase loops suffer from slug flow or dry-out when refrigerant charge is low or lines are improperly sloped. Both are detected during commissioning by monitoring individual GPU die temperatures under load.
Corrosion or galvanic attack appears when dissimilar metals are introduced at couplings without dielectric unions; this is prevented by using only the material set listed in the CDU and cold-plate specifications.
Commissioning and Leak Testing Protocols
Begin with a nitrogen hold at the pressure stated in the OEM installation manual for the hold period stated in the OEM installation manual while monitoring for decay. Follow with helium sniffing at every joint if the facility requires trace-gas certification.
Introduce coolant only after passing the gas test, then run the circulation pump at minimum speed while inspecting all connections with a calibrated moisture or refrigerant detector. Record baseline flow, pressure drop, and GPU inlet/outlet temperatures before applying full compute load.
Document every step in the commissioning package so that later service teams know the exact configuration and test results.
Decision Criteria for Rack Deployment
Choose single-phase when the facility already operates water-based CDUs and when rack power densities allow the higher flow rates required. It tolerates simpler field repairs and uses widely available maintenance staff.
Select two-phase when the design targets higher heat flux per cold plate or when the facility chilled-water temperature is too warm for single-phase rejection. The sealed nature reduces long-term fluid management but increases the skill level needed for any repair that breaks the refrigerant boundary.
In both cases the rack-level NVLink fabric remains copper; fiber and MPO infrastructure handles only the scale-out network and is routed separately from the liquid loops.
Standards referenced: ASHRAE TC 9.9 liquid cooling guidelines · OEM GPU cold-plate mounting torque specifications · Manufacturer CDU installation and leak-test procedures
Frequently asked_
How do you isolate a leaking cold plate without draining the entire rack loop?
Close the isolation valves at the manifold branch for that GPU. Use the quick-disconnect on the cold-plate side to remove the unit while the rest of the loop remains pressurized. Replace the plate, reconnect, reopen the valves, and re-bleed only that branch before returning the rack to service.
What is the first check when GPU temperatures rise after a two-phase install?
Verify liquid level at the condenser sight glass and confirm the condenser fan or pump is running. Next inspect the refrigerant line slopes for traps. Only after those checks pass do you examine individual cold-plate mounting torque and thermal interface condition.
Can single-phase and two-phase loops share the same facility headers?
No. The fluids, operating pressures, and material compatibility requirements differ. Separate headers and CDUs are required; cross-connection risks corrosion, fluid contamination, and loss of the two-phase charge.
How often should quick-disconnect seals be replaced during rack moves?
Inspect the O-ring and sealing surface on every disconnect during each rack de-install. Replace the seal if any deformation, cut, or contamination is visible. Never reuse a seal that has been under pressure for more than one year without inspection.
Who performs the helium leak test on a new two-phase rack?
The installing crew performs the test using calibrated equipment and records the final leak rate. Leviathan Systems crews include the test report as part of the rack commissioning package handed to operations before the rack is released for compute workload.