Liquid Cooling_
Manifolds & Quick-Disconnects in DLC: Field Handling Without Drips
Details field procedures for installing and servicing in-rack coolant manifolds, dripless quick-disconnects, and GPU cold plates in direct liquid cooling loops on H100 through GB300-class systems, with emphasis on sequence, inspection, and leak prevention.
Key facts
- Manifold assemblies ship pre-terminated from the OEM with fixed port spacing that must match the rack U positions of the cold plates.
- Dripless QDs require full engagement of the double-shutoff valves before any coolant path opens; partial engagement is the leading source of micro-leaks.
- Cold plates are attached to GPUs with specific thermal interface material and retention hardware whose torque sequence is defined in the GPU OEM service manual.
- Pressurized leak testing uses the facility coolant at the design operating pressure and holds for a minimum duration stated in the rack-level commissioning procedure.
- Service of a single cold plate requires isolation valves on the manifold branch or a portable service cart with bypass loops to avoid draining the entire rack.
- Particle contamination from open QDs during handling is controlled by keeping protective caps in place until the moment of connection and by using only lint-free wipes approved for the coolant chemistry.
- Leviathan Systems crews perform these steps as part of rack integration for hyperscale AI deployments.
Receiving and staging manifold assemblies
Inspect every manifold for shipping damage to the aluminum or stainless body, the mounting brackets, and the QD ports. Verify that all protective caps are present and that the port labels match the rack elevation drawing.
Record the manifold serial numbers against the bill of materials before the assembly leaves the receiving area. Any mismatch requires escalation before the unit is lifted into the rack.
Stage the manifolds on padded carts so that the QD faces remain horizontal and never rest on the sealing surfaces. This orientation prevents debris from settling into the valve seats and keeps the double-shutoff mechanisms from binding during later engagement.
Mounting the manifold to the rack frame
Position the manifold so its supply and return ports align with the planned cold-plate locations. Use the rack manufacturer’s slotted mounting holes and the supplied isolation grommets to prevent galvanic contact between dissimilar metals.
Torque the mounting fasteners in a cross pattern to the value listed on the manifold installation drawing. Over-torquing distorts the manifold bar and misaligns the QDs, which later produces side load on the nipples.
Confirm that the manifold remains level within the tolerance shown on the drawing. A slope outside tolerance traps air pockets at the highest ports and creates uneven flow distribution across the cold plates.
Connecting dripless QDs to GPU cold plates
Remove the protective cap from both halves only when the GPU tray is fully seated and the cold-plate ports are within reach. Align the QD bodies by hand and push straight until the audible click and visual indicator ring confirm full engagement of the double-shutoff valves.
Never rotate or rock the connector during mating. Side load damages the internal shutoff valves and creates a leak path that appears only after thermal cycling. After connection, tug gently on the hose to verify the latch has captured the nipple.
Repeat for every branch, working from the bottom of the rack upward so that any residual fluid drains away from already-connected ports.
Initial pressurization and leak verification
Close all drain valves and open the isolation valves at the rack inlet. Bring the loop to the design static pressure using the facility fill cart and hold for the duration specified in the commissioning checklist.
Walk the rack with a calibrated electronic leak detector or use the OEM-approved soap solution on every QD and manifold joint. Any bubble formation or detector reading above background requires immediate isolation of that branch.
Document the pressure decay curve. A stable reading confirms the loop is ready for dynamic flow testing and GPU power-on.
Cold-plate replacement during service
Isolate the affected manifold branch with the integral ball valves or install a service cart bypass. Depressurize only the local segment before breaking the QDs; the remainder of the rack stays at pressure.
Support the GPU tray on the provided service rails so the cold plate does not drop when the last fastener is released. Transfer the existing thermal interface material to a clean tray and apply new material exactly as specified in the GPU service procedure.
Reconnect the QDs, reopen the isolation valves, and repeat the static pressure hold before returning the node to the compute fabric.
Common failure modes and field detection
Partial QD engagement produces a slow weep that only appears after the system reaches operating temperature and vibration. Always verify the indicator ring is fully seated and perform a second visual check after the first thermal cycle.
Contamination trapped in the QD seal from dropped caps or dirty gloves creates a leak path that passes initial pressure testing but fails after weeks of flow. Enforce cap discipline and wipe every sealing surface immediately before mating.
Misaligned manifold mounting causes binding at the QDs that eventually cracks the nipple collar. Measure port-to-port spacing with a go/no-go gauge before any connections are made.
Standards referenced: ASHRAE TC 9.9 liquid cooling guidelines · OEM GPU cold-plate service manual torque sequences · Rack-level commissioning pressure-hold procedures
Frequently asked_
How long must the static pressure hold last before the rack is released for compute bring-up?
The duration is stated in the specific rack commissioning procedure supplied by the system integrator. Most hyperscale deployments require a minimum hold of several hours with continuous pressure logging. Any decay outside the allowed band triggers re-inspection of every QD and manifold joint before the hold is restarted.
Can a single cold plate be changed without shutting down the entire rack loop?
Yes, provided the manifold branch has isolation valves or a service cart with bypass loops is used. The rest of the rack remains at pressure and flow while the affected segment is isolated and depressurized. Re-pressurization and leak verification of only that branch is required before the node returns to service.
What indicates that a QD has not fully engaged?
The visual indicator ring must be completely flush with the body and an audible click must be heard. After connection, a light tug on the hose should meet solid resistance. If either check fails, the connector must be separated, inspected for debris, and re-mated.
Who performs the final leak sign-off on a Leviathan Systems rack integration?
The Leviathan lead technician and the customer’s commissioning engineer both witness the final pressure-hold record and sign the checklist. The signed document is required before the rack is handed over for network and GPU validation.
What tool is used to confirm manifold port alignment before QD connections?
A go/no-go spacing gauge matched to the cold-plate port pitch is placed across the manifold face. Ports outside tolerance require the manifold to be re-mounted before any hoses are connected.