Liquid Cooling_
Liquid-Cooling Loop Commissioning for GB200 & GB300 NVL72: Flush, Fill, Leak-Test, Accept
The field procedure for commissioning a direct-to-chip liquid-cooling loop on a GB200/GB300 NVL72 rack — flush, fill and air purge, pressure and leak testing, coolant acceptance criteria, and CDU startup — with the thresholds that count as a pass.
Key facts
- GB200/GB300 NVL72 racks are 100% liquid-cooled — there is no air-cooled fallback for the GPUs
- A single NVL72 rack rejects on the order of ~120 kW into the coolant loop
- Coolant is typically a propylene-glycol/water blend (often ~PG25) with controlled conductivity and pH
- The loop must be flushed and air-purged before load — trapped air causes hot spots and pump cavitation
- Acceptance gate: zero pressure decay on a timed hold and zero leak-detection alarms before energization
Why the loop is commissioned before anything powers on
On a GB200 or GB300 NVL72 rack, the GPUs are cooled by direct-to-chip cold plates fed from the rack manifold and the coolant distribution unit (CDU). There is no air-cooled mode to fall back on, so the cooling loop has to be proven clean, full, leak-free, and flowing at the design rate before any tray draws real power. Commissioning the loop is a gate, not a formality: a hidden leak or a pocket of trapped air does not announce itself until the rack is hot and under load, when it is far more expensive to find.
The work breaks into four stages — flush and fill, air purge, pressure and leak test, and CDU startup with flow verification — each with a pass criterion that has to be met before the next begins.
Flush and fill
Before the design coolant goes in, the loop is flushed to remove manufacturing debris, flux, and particulate that would otherwise lodge in cold-plate microchannels. Flush with the specified fluid, circulate, and drain until the return runs clean; particulate in a cold plate is a silent throttle on heat transfer.
Fill with the specified coolant — commonly a propylene-glycol/water blend — and confirm it meets the spec for conductivity, pH, and particulate. Off-spec coolant corrodes wetted materials and fouls plates over the life of the deployment, so this is verified, not assumed.
Purge the air
Air trapped in the manifold headers and cold plates is the most common cause of post-commissioning hot spots and pump cavitation. Bleed and vent every high point as the loop fills, then circulate and re-purge until the system holds full flow without entrained air. On a tall NVL72 manifold the top trays are the usual offenders — they get verified individually.
Pressure and leak test
With the loop full and purged, pressurize to the OEM test criteria and hold. The acceptance gate is zero pressure decay over the timed hold and zero alarms from the rack's leak-detection system. Every quick-disconnect between the manifold and each compute or switch tray is checked for full seat and no drips — a QD that is not fully home will pass a glance and fail under thermal cycling.
Leak detection is commissioned in the same pass: confirm the sensors are placed, powered, and actually trip, rather than assuming they will.
CDU startup and flow verification
Bring the CDU up per the OEM Method of Procedure: verify supply and return temperatures sit within spec, that flow rate meets the design value for the full rack heat load, and that pump redundancy is real — pull-test a pump and confirm the standby carries the loop. Record the supply temperature, flow, and pressure drop as the commissioning baseline; these are the numbers the operations team will monitor against for the life of the cluster.
Only once the loop is clean, full, leak-free, flowing at rate, and baselined does the rack move on to power-on, POST, and the rest of the acceptance test plan.
Standards referenced: ASHRAE liquid-cooling guidance · OEM CDU / cold-plate specifications · Project acceptance test plan (ATP)
Frequently asked_
How do you commission the liquid-cooling loop on a GB200 or GB300 NVL72 rack?
In four gated stages: (1) flush the loop and fill with spec coolant, verifying conductivity, pH, and particulate; (2) purge all trapped air from manifolds and cold plates; (3) pressure and leak test to OEM criteria — the gate is zero pressure decay on a timed hold and zero leak-detection alarms, with every quick-disconnect verified seated; (4) start the CDU, confirm flow rate, supply/return temperatures, and pump redundancy, and record the commissioning baseline. Only then does the rack power on.
What is the acceptance criteria for a GPU liquid-cooling loop?
Zero pressure decay over the timed pressure hold, zero leak-detection alarms, every quick-disconnect fully seated with no drips, coolant within spec for conductivity and pH, and the CDU delivering design flow at the correct supply/return temperatures with verified pump redundancy — all captured in the as-built commissioning baseline.
Who commissions liquid cooling for GPU clusters?
Leviathan Systems commissions direct-to-chip liquid-cooling loops on GB200 and GB300 NVL72 deployments across the United States — flush and fill, air purge, pressure and leak testing, CDU startup, and flow verification — delivering the commissioning baseline and acceptance sign-off as part of the as-built package.