What happens in the first 30 seconds of a pump trip
A liquid-cooled AI rack carries most of its heat away in the coolant. A GB300 NVL72 puts roughly 120 kW into the loop. When the CDU pump trips, that heat keeps coming while the flow dies away, and the only thing between the GPUs and their thermal limit is the heat the cold plates and the coolant already inside the rack can soak up.
Two numbers matter. Ride-through is how long the rack lasts if flow never comes back. Latest standby start is the longest delay the standby pump can take to start and still keep the cold plates under their limit. If your control sequence detects the trip, decides, and starts the standby pump slower than that, the rack throttles or shuts down.
Both depend on margin before the trip. A rack that runs at its vendor's hottest allowed point has close to zero ride-through. Every degree of colder supply, or every extra litre per minute, buys seconds.
The model
Each rack is two thermal masses joined by a resistance: the cold plates and silicon (Cb) and the coolant inside the rack (Cf). The tripped pump coasts down with torque proportional to speed squared, and flow follows speed:
The limit is the cold-plate temperature at the vendor's hottest sanctioned operating point: the outlet limit plus the base-to-coolant rise at rated flow. The standby pump starts after the delay and ramps linearly; with check valves, the faster of the two pumps sets the flow. The equations are integrated with fourth-order Runge-Kutta at 50 ms steps.
We checked the calculator against a full network model of a 16-rack GB300 hall on a 2.3 MW CDU, with every pipe, check valve and the CDU heat exchanger. For the worst rack the calculator gives 17.8 s of ride-through against 17.6 s from the network model, and a latest standby start of 14.3 s against 14.0 s.
Rack data used
| Rack | Heat to liquid | Rated flow | Outlet limit | Source |
|---|---|---|---|---|
| GB300 NVL72 | 121.5 kW | 177 L/min at 45 °C | 55.4 °C | Lenovo LP2357 (90% of 135 kW to liquid); minimum flow vs supply temperature from Lenovo's environmental table: 59, 71, 89, 119, 177 L/min at 25, 30, 35, 40, 45 °C. The outlet limit is derived: every row of that table lands at 55.4 to 56.1 °C. |
| GB200 NVL72 | 115 kW | 130 L/min | 65 °C | QCT QoolRack paper: supply up to 45 °C, return up to 65 °C. |
| Generic cold-plate rack | 100 kW | 150 L/min | base 80 °C | OCP OAI liquid cooling guideline: 1.25 to 2.0 L/min per kW, typical 1.5. |
Cold-plate rise (24 K) comes from OCP's published case-to-inlet resistance of 0.014 to 0.019 °C/W at about 1.4 kW per GPU. Cold-plate heat capacity and rack coolant volume are estimates; change them under "Rack thermal properties" if you have vendor numbers.
Coolant density and specific heat are taken at 40 °C: treated water from standard water tables, propylene and ethylene glycol mixtures from Engineering ToolBox tables. Propylene glycol at 25% by volume (PG25) is the Open Compute Project standard for rack loops. Ethylene glycol carries heat slightly better but is toxic, so it is rarely used inside rack loops where quick disconnects drip during swaps.
Questions
How long can a GB300 NVL72 rack survive loss of coolant flow?
It depends on how close to its limit it runs. At 35 °C supply and 140 L/min of PG25, about 17.5 seconds with a pump that coasts down in under a second. At 45 °C supply and the vendor minimum of 177 L/min, about 2 seconds: the cold plates start within a fraction of a degree of their limit.
How fast does the standby pump need to start?
Faster than the ride-through time, because a ramping pump does not restore full flow instantly. At 40 °C supply and 150 L/min per GB300 rack, the standby pump has to start within about 7 seconds. An 8 second start delay takes the cold plates over the limit.
What does the coast-down time constant mean?
It is the time for the tripped pump to fall to half speed. Heavy rotors and low shaft power coast longer. A pump with 0.15 kg·m² of inertia at 1800 rpm and 7.4 kW of shaft power has τ ≈ 0.7 s. A flywheel or a pump on a UPS raises it a lot; try it in the calculator.
Why does a colder supply temperature help so much?
Ride-through is mostly margin divided by heating rate. Once the flow is gone, the cold plates heat at power over heat capacity, about 0.75 K/s for a GB300 rack, whatever the supply temperature. Every degree of colder supply adds a degree of margin, which buys about 1.5 seconds.
Is this a substitute for a design review?
No. It is a first-order estimate for one rack with the supply held steady. Use it to see whether a control sequence is in the right range, then check the full loop.