Installation_
De-Racking & Decommissioning a GPU Cluster for Migration
This article provides the exact sequence for de-racking H100 and later NVL72 clusters, covering power isolation, copper NVLink spine handling, fiber scale-out disconnection, asset tagging, and crate packing so crews can relocate racks without damage or documentation loss.
Key facts
- NVLink connections inside NVL72 racks use copper backplanes and do not involve MPO or fiber.
- Scale-out InfiniBand or Ethernet links run on separate MPO trunk cables between racks and switches.
- Lockout/tagout follows OSHA 1910.147 before any power work on PDUs or busways.
- Factory-terminated MPO trunks must be inspected with a calibrated MPO continuity tester or microscope before storage.
- Racks follow EIA-310-D mounting dimensions and require sequential unbolting from top U positions downward.
- ESD grounding straps must remain attached to the rack frame during card removal to prevent latent damage.
- Asset tags must record serial numbers, rack U position, and cable endpoints before any disconnection.
Pre-teardown inventory and labeling
Begin with a full rack audit that records every GPU tray serial, switch UID, and PDU identifier along with its exact U position. Photograph each cable bundle at both ends and assign unique labels that stay with the cable during removal. This record becomes the only reliable map when racks are reassembled at the destination site.
Two technicians cross-check each entry to eliminate single-point transcription errors. Any mismatch between the physical label and the database halts work until resolved.
Power isolation and safe shutdown
Apply lockout/tagout devices to all branch circuits feeding the rack PDUs before touching any power cords. Verify zero voltage at the GPU tray inputs with a calibrated meter after the servers report shutdown complete. This order prevents inrush damage to power supplies and eliminates arc-flash exposure during cord removal.
Confirm that the NVLink copper spine remains powered only through the trays themselves; once trays are off, the backplane carries no voltage. Never rely on remote management alone for final verification.
Scale-out fiber and NVLink disconnection
Disconnect MPO trunks carrying the InfiniBand or Ethernet fabric first, capping each connector immediately and routing the cable to a dedicated storage reel. Leave the internal copper NVLink backplane untouched until the trays are fully extracted; it is not field-serviceable and must stay with the rack frame.
Clean and inspect every MPO endface with a microscope before coiling. Dust trapped during transport is the most common cause of link failures after re-installation.
Tray extraction and rack disassembly
Remove trays starting at the top U and working downward so the rack center of gravity stays low. Support each tray on a two-person lift or mechanical assist until it clears the rails; GPU trays exceed safe single-person weight limits. Store trays in anti-static bags inside the same crate they will travel in.
Unbolt the rack from the floor and overhead bracing only after all trays and switches are out. Keep the copper NVLink spine and backplane assembly inside the rack during transport to avoid misalignment that would require factory recalibration.
Common failure modes during decommissioning
The most frequent error is pulling an MPO trunk before the connected switch port is administratively shut down, which triggers link-flap storms on adjacent racks still in production. Always coordinate with the NOC to drain traffic first.
A second pattern is mismatched labeling that forces technicians to re-trace cables after the rack is already crated. This occurs when labels are applied only at one end or when two technicians use different naming conventions. The third recurring issue is bending the copper NVLink cables that exit the backplane during tray removal; even minor kinks create intermittent training failures later. Catch these by requiring a second visual inspection of every cable path before the tray leaves the rails.
Packing, transport, and documentation handover
Crate each tray with its original foam inserts and desiccant packs, then seal the crate with serialized tamper-evident tape. Record crate ID against the tray serial list before the truck leaves the loading dock. The same list travels with the shipment and is reconciled on arrival before any uncrating begins.
Hand the complete digital and paper asset package to the receiving team at the same time the last crate is loaded. Missing documentation has caused entire racks to sit idle for weeks while serial numbers are re-collected.
Standards referenced: OSHA 1910.147 · EIA-310-D · TIA-942
Frequently asked_
Do we need to break the NVLink domain when moving an entire NVL72 rack?
No. The copper NVLink spine stays inside the rack frame and travels with it. Only the scale-out MPO trunks that connect to the rest of the fabric are disconnected. Reconnecting those trunks at the new location restores the external network; the internal NVLink topology remains intact.
What test equipment travels with a de-rack crew?
A calibrated MPO continuity tester, fiber microscope, and non-contact voltage detector are minimum kit. The crew also carries the original OEM torque driver for any PDU or busway lugs that must be re-torqued on the destination side. No field termination tools are required because all MPO trunks are factory-terminated.
How long should power be left off before tray removal?
Follow the OEM cool-down period listed in the tray service manual, typically measured in minutes rather than hours. The goal is to let capacitors discharge and fans stop, not to reach ambient temperature. Voltage verification with a meter is still required regardless of elapsed time.
Who owns the final asset list after migration?
The receiving site signs the reconciled list before any crates are opened. A copy is retained only for warranty and transport-damage claims. Any discrepancy discovered later is logged against the signed document.