Zinc Whiskers Under Raised Floor Caused Failure of 128-Rack AI Cluster
At a 2 MW AI data center with 128 racks of liquid-cooled servers running eight Nvidia H100 GPUs each, zinc whiskers from galvanized floor supports entered servers and triggered short circuits, ECC errors and GPU resets. Recovery required board replacements, particle filters and replacing zinc-plated components.
- Cluster: 128 racks, eight 700 W Nvidia H100 GPUs per server
- Whiskers 0.5–2 mm long and 1–5 µm thick bridged traces on GPU boards
- Failures peaked during training when fan speeds and power draw were highest
- Fix: board replacement, particle filters and non-zinc floor components
Read next
Hardware