1. Spot it: a monitoring alert or diagnostic failure flags a bad card.
2. Isolate it: pull the node out of the job queue.
3. Confirm it: run habana-validate to make sure it's really broken.
4. File the claim: through Intel's Partner Support portal. Note — this can take longer than it would for an actively-promoted product, so we set expectations with the customer up front rather than promising standard turnaround.
5. Swap in a spare: from our on-site cold spare pool.
6. Bring it back online: validate the replacement, confirm the driver loads cleanly.

Every card that leaves the building — whether for RMA, retirement, or end of contract — gets wiped first (per NIST data-sanitization standards) and logged with serial number, timestamp, and who did it.

← Previous | Guide Index | Next →