LEVIATHAN SYSTEMS

Field Guide_

GB200 NVL72 Installation Checklist_

A field checklist for installing a single NVIDIA GB200 NVL72 rack — from site readiness through commissioning and hand-off. One NVL72 rack lands roughly 120 kW of power, a ~1,360 kg chassis, a mandatory liquid-cooling loop, and a cable plant on the order of 2,000 connections. This is the installer's sequence, with acceptance criteria and the standards each step answers to.

01

Site & Rack Readiness (before the rack ships)

ASHRAE TC 9.9 · TIA-942

  • Confirm power: ~120 kW per rack; verify PDU/busway feed, breaker sizing, and redundancy (2N / N+1) against the rack one-line.
  • Verify floor loading: a populated NVL72 is ~1,360 kg (3,000 lb). Confirm tile/slab rating and pathway to the rack location.
  • Cooling: facility water loop available at design ΔT and flow; CDU sized for full rack heat (≈120 kW); supply water temperature within spec.
  • Network: leaf/spine switches racked and powered; fiber pathways and ladder rack to the GPU rack staged.
  • Receiving: inspect crates for impact/tilt indicators; stage in a clean, ESD-controlled area at room temperature before unboxing.
02

Rack Landing & Mechanical

TIA-942 · OEM MOP

  • Position and level the rack on its destination tile; engage seismic/anchor hardware per the facility AHJ.
  • Install compute trays and NVLink switch trays per the OEM slot map; verify blind-mate alignment and tray seating.
  • Torque all mechanical fasteners to spec; confirm no shipping brackets remain.
  • Bond the rack to the data center grounding system; verify ground continuity (<1 Ω to the common bonding network).
03

Power

NEC · IEC 60950 / 62368

  • Land power whips to PDUs; verify phase balance across the rack and per-PDU load headroom.
  • Confirm bus bar / power shelf seating on each tray; check for correct polarity and locking.
  • Power-on sequence per OEM MOP; verify all PSUs report healthy and redundancy is real (pull-test one feed).
  • Record per-phase current at idle and note expected ramp under load for the commissioning baseline.
04

Liquid Cooling (mandatory on NVL72)

ASHRAE liquid cooling · OEM CDU spec

  • Flush and fill the loop with spec coolant; purge air from manifolds and cold plates.
  • Connect quick-disconnects to every compute and switch tray; verify full seat and no drips.
  • Pressure / leak test the rack loop to OEM criteria; hold and verify zero pressure decay.
  • Commission leak detection; confirm CDU flow rate, supply/return temperatures, and pump redundancy.
05

The Cable Plant (~2,000 connections per rack)

TIA-606 labeling · IEC 61300-3-35 inspection · BICSI

  • NVLink spine: connect the 72-GPU NVLink domain per the OEM map — every cartridge fully seated, correct port pairing.
  • Back-end fabric: land InfiniBand (NDR/XDR) or Spectrum-X Ethernet to leaf switches; maintain MPO polarity and bend-radius minimums.
  • Inspect every fiber endface to IEC 61300-3-35 before insertion; clean-inspect-connect, no exceptions.
  • Label both ends of every cable to a TIA-606 scheme; build the as-built cable map as you go, not after.
  • Dress and strain-relief; preserve service loops and airflow/serviceability clearances.
06

Test, Commission & Hand-Off

Acceptance test plan (ATP)

  • Fiber: OTDR / insertion-loss + return-loss test every link; record results against pass thresholds.
  • Power-on self-test (POST) on all trays; resolve any GPU/NVLink/port faults before fabric bring-up.
  • Fabric validation: link-up on every port, no flapping; NCCL all-reduce / bandwidth test across the domain.
  • Thermal soak / burn-in under load; verify no thermal trips and stable temperatures.
  • Deliver the as-built package: cable map, test reports, ATP sign-off, photos, and open-items list.

Acceptance Criteria_

ItemPass Criteria
Fiber endfacePass IEC 61300-3-35 zones (clean) before every insertion
Insertion lossWithin channel budget for OM4/OM5 / OS2 link type
Return lossMeets connector-grade threshold (APC/UPC as specified)
Liquid loopZero pressure decay on hold; zero leak-detection alarms
PowerAll PSUs healthy; redundancy verified by live feed pull-test
Fabric100% ports link-up, no flapping; NCCL bandwidth within expected range
ThermalNo thermal trips through burn-in; temps stable under sustained load
DocumentationTIA-606 labels both ends; complete as-built cable map + ATP sign-off

This checklist is a field reference, not a substitute for the OEM Method of Procedure (MOP) or facility-specific acceptance test plan. Leviathan executes this layer end-to-end on live GB200 and GB300 deployments.

Ready to Deploy Your GPU Infrastructure?_

Tell us about your project. Book a call and we’ll discuss scope, timeline, and the best approach for your deployment.

Book a Call