Installation_
GPU Rack Receiving, Staging & Lift Plan: Moving ~1,360 kg Racks Without Damage
A field-tested, step-by-step guide for receiving, staging, and rigging heavy NVL72 GPU racks (~1,360 kg check OEM spec) into AI data centers, covering inspection checkpoints, staging layout, lift planning, and common damage modes—written for deployment engineers who move these racks daily.
Key facts
- An NVL72 rack with 72 GPUs, liquid cooling, and copper NVLink backplane typically weighs approximately 1,360 kg (3,000 lb) empty of coolant, and up to around 1,600 kg when filled with coolant and networking gear—verify exact weight per OEM spec.
- Rack center-of-gravity (CoG) is high—often above the midpoint due to GPU trays and power supplies; lift plans must account for this to prevent tipping during transport.
- Standard data-center raised-floor tiles are commonly rated for about 1,200 kg point load; a loaded NVL72 rack exceeds this, so a load-spreading plate (e.g., 1/2-inch steel plate, 1.2 m x 1.2 m) is required under the rack during staging and installation.
- The TIA-942-B standard specifies minimum aisle width of 1.2 m for hot/cold aisles; for rack movement, a 1.5 m clear path is recommended to allow pallet jack or lift truck maneuvering without scraping columns or adjacent racks.
- A calibrated MPO continuity tester (e.g., a 12-fiber MPO loopback) is used to verify fiber trunk cables after routing, but NVLink copper backplane connections are tested via nvidia-smi 'nvlink' command—never via fiber tools.
- Rack-level vibration during transport should not exceed the OEM shipping spec (typically 0.5 g RMS in any axis); exceeding this can loosen GPU tray screws or damage liquid cooling quick-disconnects.
- The rack's shipping pallet must be removed only after the rack is on the load-spreading plate and leveled; premature removal can cause the rack to tip forward due to high CoG.
Receiving Inspection: What to Check Before the Truck Leaves
Before the delivery truck departs, perform a walk-around inspection with the driver present. Document any visible damage to the outer crate or rack frame—especially dents, scratches, or bent rails—using photos and a signed damage report. Check the shipping label for tilt indicators and shockwatch tags; if any are tripped (e.g., tilt beyond the OEM's specified limit, typically 30 degrees, or shock above the OEM threshold, usually 2 g), reject the rack or note it for OEM warranty claim. Inside the crate, verify that all foam blocks and corner braces are intact; missing or shifted foam indicates rough handling that may have stressed the copper NVLink backplane or liquid cooling manifolds.
Use a calibrated digital level to measure rack tilt on the pallet: the rack should be within 1 degree of level in both axes per the OEM installation manual. If the rack is tilted more than 2 degrees, the liquid cooling loop may have air pockets that require purging before power-on. Also confirm that the pallet is rated for the rack weight (typically a 2,000 kg capacity pallet with four-way fork entry; check the pallet rating plate). Do not accept a rack on a damaged pallet—it can collapse during staging.
Staging Layout: Floor Loading, Clearances, and Sequencing
Stage racks in a dedicated area with a concrete floor that meets the data center's load specification for point loads (check the structural design documents). Place a load-spreading plate (e.g., 1/2-inch steel plate, 1.2 m x 1.2 m) under each rack position to distribute the concentrated weight over a larger area, preventing tile cracking or floor depression. Maintain a minimum 1.5 m clear path between staged racks and walls or columns for pallet jack and lift truck access. Sequence racks by row and position per the deployment plan: stage the first row closest to the installation aisle, then the second row behind it, to minimize back-and-forth movement.
Allow at least 0.5 m gap between adjacent staged racks for access to side panels and cable management. If racks are liquid-cooled, stage them near the coolant distribution unit (CDU) location to reduce hose runs. Keep the staging area clean of debris and moisture; a single dropped screw can puncture a coolant hose during movement. Use floor markings (tape or paint) to define rack footprints and aisle boundaries.
Lift Plan: Rigging, Center of Gravity, and Maneuvering
The lift plan must account for the rack's high center of gravity (CoG), typically 60-70% of the rack height from the floor due to GPU trays and power supplies. Use a lift truck or pallet jack with a rated capacity of at least 2,000 kg and a low-profile fork (e.g., 50 mm fork height) to slide under the rack pallet. If the rack is on a skid without a pallet, use a spreader bar and slings rated for 2,500 kg with a safety factor of 5:1, attached to the rack's lifting brackets (never the frame rails). Always lift from the bottom—never tilt the rack more than 10 degrees from vertical during movement, as the CoG can shift and cause tipping.
When moving the rack into the data center aisle, use a guided path with spotters at front and rear. For raised-floor environments, place a temporary plywood or steel bridge over the floor tiles to distribute the load while the rack is wheeled into position. Once at the final location, lower the rack onto the load-spreading plate and level it using the rack's leveling feet (typically M12 threaded feet with a 25 mm adjustment range). Verify level with a digital level: within 0.5 degrees in both axes. Do not remove the shipping pallet until the rack is on the plate and leveled, as the pallet provides stability against tipping.
Common Failure Modes: What Goes Wrong and How to Catch It
The most common failure is rack tipping during unloading or staging. This occurs when the pallet is removed prematurely or the rack is tilted beyond 10 degrees while on a pallet jack. To catch it, always keep the rack vertical and use a tilt indicator on the crate; if the indicator shows tilt beyond the OEM's transit limit (typically 30 degrees) during transit, inspect the rack for internal damage—especially the copper NVLink backplane, which can crack at the solder joints. Another frequent issue is floor tile collapse under the rack weight. Always use a load-spreading plate; a standard raised-floor tile (rated ~1,200 kg) will crack under a 1,360 kg rack, causing the rack to sink and misalign with adjacent racks.
Liquid cooling quick-disconnect damage is also common: if the rack is jostled during movement, the QDs can leak or break. Inspect all QDs for cracks or leaks after staging. Finally, fiber trunk cables (MPO) can be pinched or bent beyond the minimum bend radius (typically 10x the cable diameter) if routed too tightly during rack placement. Use a bend-radius gauge to check all fiber runs after the rack is seated. If any failure is found, stop the deployment and document it for the OEM—do not attempt field repair of copper backplane or liquid cooling loops.
Copper NVLink vs. Fiber Network: Separate Domains, Separate Testing
In NVL72-class racks, the GPU-to-GPU NVLink connections run over the copper NVLink spine/backplane inside the rack—not over fiber or MPO. This copper backplane is factory-terminated and tested; field work is limited to verifying that the GPU trays are fully seated and that the backplane connectors are clean. Use nvidia-smi 'nvlink' command to check link status (e.g., 'nvidia-smi nvlink -s' shows link counts and errors). Do not confuse this with the scale-out compute network (InfiniBand or Ethernet), which uses MPO/fiber trunks between racks and switches. The fiber network is a separate domain: MPO trunk cables are factory-terminated and polished; field work is patching, routing, cleaning, inspection, and testing with a calibrated MPO continuity tester or OTDR.
Never imply that 'NVLink up' depends on a fiber/MPO link—they are independent. If an NVLink link is down, it is almost always a seating issue or a damaged backplane, not a fiber problem. Conversely, if the compute network is down, check the MPO connections and switch ports. This distinction is critical for troubleshooting: chasing a fiber issue for an NVLink failure wastes hours.
Liquid Cooling Considerations: Weight, Leaks, and Purging
Liquid-cooled NVL72 racks are heavier than air-cooled equivalents due to coolant (typically deionized water or dielectric fluid) in the manifolds and cold plates. The rack weight can increase by 200-300 kg when filled—check the OEM specification for exact values. During staging, keep the rack empty of coolant until it is in its final position and leveled; this reduces the risk of leaks during movement. After placement, fill the cooling loop using the CDU, following the OEM's fill-and-purge procedure. Use a pressure gauge to verify the loop is at the specified operating pressure (typically 30-50 psi per OEM) and check for leaks at all QDs and hose connections with a leak detection fluid (e.g., soapy water or electronic leak detector).
If a leak is detected, isolate the affected section and replace the QD or hose—never attempt a field repair on a cold plate. After filling, purge air from the loop by running the CDU pump at low speed for the duration specified by the OEM (typically 15 minutes) while monitoring flow rate. Air pockets can cause GPU overheating; verify that all GPU cold plates are within 2°C of each other using nvidia-smi 'temperature' command. Document the fill pressure and temperature for the commissioning report.
Commissioning Sequence: Power, Network, and NVLink Verification
After the rack is leveled, filled with coolant, and connected to power and network, follow this sequence: First, apply facility power (typically 480 VAC 3-phase) and verify the rack's power distribution unit (PDU) is receiving voltage within ±10% of nominal. Second, power on the rack's management controller (BMC) and check for any hardware alerts (e.g., GPU tray not seated, fan failure). Third, bring up the compute network by connecting the MPO trunk cables from the rack's top-of-rack switch to the leaf switches, then test with a calibrated MPO continuity tester to ensure all fibers are passing. Fourth, run nvidia-smi 'nvlink -s' to confirm all NVLink links per GPU are active (the number varies by GPU model: 18 for H100/H200, 36 for GB300; check the GPU specification). If any link is down, reseat the GPU tray and clean the backplane connector with a lint-free wipe and isopropyl alcohol.
Finally, run a stress test (e.g., NCCL all-reduce benchmark) to validate GPU-to-GPU bandwidth and network latency. Document all test results in the commissioning report. If any step fails, stop and troubleshoot before proceeding—a single bad link can degrade cluster performance by 10-20%. Leviathan Systems follows this exact sequence on every deployment to ensure zero rework.
Standards referenced: TIA-942-B (Telecommunications Infrastructure Standard for Data Centers) · ANSI/BICSI 002-2019 (Data Center Design and Implementation Best Practices) · OEM shipping spec for NVL72 rack (vibration < 0.5 g RMS, tilt < 30 degrees, check OEM documentation) · ISO 4406 (fluid cleanliness for liquid cooling loops, typically ISO 18/16/13)
Frequently asked_
What is the maximum tilt angle allowed when moving an NVL72 rack on a pallet jack?
The rack should never be tilted more than 10 degrees from vertical during movement. The high center of gravity (CoG) makes it prone to tipping beyond this angle. Use a tilt indicator on the crate to monitor transit tilt; if the indicator shows >30 degrees (or the OEM limit), inspect the rack for internal damage before staging. Always keep the rack vertical and use spotters during maneuvering.
Do I need a load-spreading plate for every rack, even on a concrete floor?
Yes, even on a concrete floor, a load-spreading plate is recommended to distribute the ~1,360 kg point load over a larger area. Concrete floors can crack under concentrated loads, especially if the floor is not reinforced. On raised floors, it is mandatory because standard tiles are rated for about 1,200 kg. Use a 1/2-inch steel plate at least 1.2 m x 1.2 m under each rack.
How do I test the copper NVLink backplane after rack movement?
Use the nvidia-smi 'nvlink' command (e.g., 'nvidia-smi nvlink -s') to check link status. All links should show 'Active' with zero errors. If a link is down, reseat the GPU tray and clean the backplane connector with a lint-free wipe and isopropyl alcohol. Do not use fiber testers on the copper backplane—they are separate domains. Leviathan Systems verifies NVLink links as part of every rack commissioning.
What should I do if a liquid cooling quick-disconnect leaks during staging?
Immediately isolate the affected section by closing the nearest shutoff valve or disconnecting the hose at the CDU. Do not attempt to repair the QD in the field—replace it with a new one from the OEM kit. Use a leak detection fluid to verify the replacement is sealed. Document the leak in the commissioning report and check for air pockets in the loop after refilling.
Can I field-terminate MPO connectors for the compute network?
No. MPO trunk cables are factory-terminated and polished. Field work is limited to patching, routing, cleaning, inspection, and testing with a calibrated MPO continuity tester or OTDR. Field termination of MPO ferrules is not recommended because it requires specialized polishing equipment and can introduce high insertion loss or back reflection. Always use pre-terminated cables from the OEM or a certified vendor.