Platforms_
Air-Cooled vs Liquid-Cooled GPU Platforms: Picking the Variant for Your Site
This article explains how to measure a site's existing CRAC/CRAH and facility water capacity against GPU rack heat loads to choose air-cooled versus liquid-cooled platforms for H100 through GB300 NVL72 deployments, including rack-level integration steps performed by field crews.
Key facts
- ASHRAE TC 9.9 provides inlet air temperature and humidity envelopes that air-cooled GPU racks must stay inside to avoid throttling.
- Liquid-cooled GPU racks require a facility water loop sized to the CDU heat exchanger rating and connected through quick-disconnect manifolds.
- MPO trunk cables carry the scale-out InfiniBand or Ethernet fabric between racks and are separate from any internal copper NVLink backplane.
- Factory-terminated MPO connectors must be inspected with a calibrated scope and cleaned before mating to avoid insertion loss that affects switch-to-switch links.
- Leviathan Systems performs pre-deployment cooling audits that compare measured supply-air volume and chilled-water delta-T against OEM rack power specifications.
- Air-cooled racks route exhaust through hot aisles or chimney cabinets while liquid-cooled racks add rear-door heat exchangers or direct-to-chip cold plates fed from overhead or underfloor manifolds.
- Manifold connections and rack grounding bonds follow torque values listed in the OEM installation manual and are verified with a calibrated torque wrench.
Evaluating Existing Cooling Plant Capacity
Record the nameplate capacity and current operating load of every CRAC or CRAH unit serving the target hall. Measure supply-air temperature and volume at the perforated tile or overhead diffuser nearest the planned rack row, then compare those values to the total kW the new GPU racks will draw.
Review the chilled-water loop next: note the available flow rate at the CDU connection points and the design delta-T the plant can sustain. If the measured water-side capacity falls short of the projected rack load, air cooling becomes impractical and liquid cooling must be selected.
Document both air-side and water-side readings on the same spreadsheet so the decision matrix later can reference actual field data rather than design drawings.
Limits of Air-Cooled GPU Rack Configurations
Air-cooled racks rely on the existing CRAC/CRAH infrastructure and hot-aisle containment to keep inlet air within the ASHRAE TC 9.9 recommended envelope. When rack power exceeds what the supply-air volume can remove, inlet temperatures rise and GPUs throttle.
Field crews verify that exhaust paths remain unobstructed and that blanking panels are installed in every unused rack unit. Any shortfall in containment forces hot air recirculation and immediately reduces the safe operating envelope.
Because air cooling adds no manifolds or CDUs, deployment time is shorter, yet the site must already possess sufficient chilled-air capacity or the platform choice is invalid.
Requirements for Liquid-Cooled GPU Racks
Liquid-cooled racks need a dedicated facility water loop terminated at each row with isolation valves and pressure gauges. The CDU must be sized to the aggregate GPU heat load and must maintain the required flow and temperature differential specified by the rack OEM.
Manifolds are mounted at the top or rear of the rack and connected with quick-disconnect fittings; each connection is torqued to the value listed in the OEM manual and leak-tested with a calibrated pressure gauge before GPUs are powered.
Because the internal NVLink domain remains copper, the only fiber work is the separate MPO scale-out fabric; crews therefore route and test MPO trunks after the liquid manifolds are secured so that fiber routing does not interfere with manifold service loops.
Decision Process for Platform Selection
Plot measured air-side capacity against projected rack load first. If the air infrastructure can sustain the load while staying inside ASHRAE limits, air cooling is the lower-risk choice. When air capacity is insufficient, compare the chilled-water loop capacity against the CDU rating; only then is liquid cooling justified.
Leviathan Systems crews present both options with the measured numbers and the required plant modifications so operators can approve the platform before rack delivery.
The final sign-off includes confirmation that any new CDU or manifold work will not reduce redundancy in the existing cooling plant.
Common Failure Modes in Cooling Platform Deployment
The most frequent error is under-measuring chilled-water flow before ordering liquid-cooled racks, resulting in CDUs that cannot reach design delta-T and GPUs that thermal-throttle under sustained load. This is caught by comparing flow-meter readings against the CDU specification sheet before the purchase order is released.
A second failure occurs when manifold connections are left at hand-tight rather than torqued; vibration during commissioning then produces drips that short power distribution units. Always use a calibrated torque wrench and perform a static pressure test at the value listed in the OEM manual.
A third mode appears when air-cooled racks are installed without verifying hot-aisle containment integrity; exhaust air leaks into the cold aisle and inlet temperatures climb even though the CRAC units are within nameplate limits. Smoke-pencil testing during commissioning reveals these paths before GPUs are installed.
Coordination with Networking and Power Infrastructure
Once the cooling platform is chosen, route power whips and MPO trunks so they do not cross manifold service loops. MPO trunks are dressed along the top of the rack or in overhead trays after the liquid manifolds are pressure-tested.
Grounding bonds between rack frames, manifolds, and the facility ground grid must be verified with a low-resistance ohmmeter before any GPU is seated. This step prevents ground loops that appear only after liquid cooling pumps are energized.
Final commissioning records both cooling-loop parameters and MPO link loss so that any later performance issue can be isolated to either the thermal or the network domain.
Standards referenced: ASHRAE TC 9.9 · OEM rack installation torque specifications · TIA-942
Frequently asked_
How do I confirm my chilled-water loop can support liquid-cooled racks before ordering hardware?
Measure flow rate and supply/return temperatures at the planned CDU tie-in points during peak facility load. Compare those values to the CDU heat-exchanger rating in the OEM documentation. If either flow or delta-T is marginal, engage the facility engineer to increase pump speed or add a parallel loop before racks arrive.
What test must be completed on every manifold connection before GPUs are powered?
Perform a static pressure test at the value listed in the OEM manual using a calibrated gauge. Any drop indicates a fitting that must be re-torqued or replaced. Only after the test passes are GPUs installed and powered.
Does the internal NVLink fabric affect the choice between air and liquid cooling?
No. NVLink runs over the copper backplane inside the rack regardless of cooling method. The cooling decision is driven solely by facility air or water capacity, while MPO trunks handle the separate scale-out network.
When should Leviathan Systems be brought in for a cooling assessment?
Schedule the assessment after the site has provided as-built CRAC and chilled-water drawings but before the GPU rack purchase order. The crew records actual operating values and produces a one-page matrix that shows which platform fits the measured capacity.
What single measurement most often changes an air-cooled plan to liquid-cooled?
Supply-air volume at the rack inlet under full CRAC fan speed. When this volume cannot remove the projected rack heat load while staying inside the ASHRAE envelope, the platform must switch to liquid cooling.