Buying used GPUs for local inference is common. Skipping burn-in is how mystery artifacts and random OOMs enter production folklore.
Routine
- Visual inspection and photo inventory.
- Stock stress test for thermals and power.
- Install your actual inference stack and run a long generation at target context.
- Repeat after a cold boot; watch for post-sleep failures on desktops.
- Keep the card in a burn-in machine for several workdays before trusting customer demos.
Decision gates
Fail the unit on unstable clocks, sensor nonsense, or inference errors reproducible across runtimes. Cross-check capacity claims with match/vram-table. Build context from diy and hardware. Troubleshooting playbooks: tutorials/troubleshooting. Budget framing: budget-3000.
Extra
Document serials and seller terms. Do not mix unverified used cards into a shared queue. If a card passes games but fails long LLM sessions, believe the LLM test—the load pattern differs. Community threads help but do not replace your logs (community). For procurement help: contact.
Burn-in is dull. Mystery downtime is worse.
Warranty and finance
Align burn-in duration with return windows when possible. If the window is short, prioritize inference tests on day one. Track failure rates by seller for future purchases. Do not deploy a used card to a customer-facing demo the same day it arrives. Keep anti-static handling notes in the receiving SOP.
Test corpus for burn-in
Keep a fixed set of prompts: short chat, long summary, JSON extract, and a multi-minute generation. Record tokens per second at start and after twenty minutes to detect throttling. Save logs with timestamps. If artifacts appear only on one runtime, try a second runtime before RMA. Photograph the card serial beside the burn-in machine. Do not interrupt burn-in for opportunistic gaming tests that change thermal history. After passing, flash a sticker with the pass date on the box. Update asset management. Share anonymized failure modes with the purchasing channel so the company learns.
Integration with purchasing
Require burn-in sign-off before accounting marks the asset as deployed. Used cards that fail should re-enter seller return flows with evidence attached. Keep a spare known-good card for swapping during diagnosis. Teach receiving staff not to power used GPUs on cardboard antistatic bags near liquids. After a pass, update the internal wiki with the driver version used during burn-in. If two cards from the same seller fail similarly, pause purchasing and investigate. Share sanitized metrics monthly.
Published by MicroAI One (本地AI 1号网), a Shenzhen team working on local LLM hardware, software and deployment.