How to inspect a used GPU server

By the RentServers.com team · Updated October 28, 2026

Before you buy: paperwork and serials

• Serial numbers for the chassis, the GPU baseboard and each GPU. Ask for photos of the labels and match them to the seller's list.
• Warranty status: check every serial with the original manufacturer. Ask whether any warranty or support contract transfers.
• History: where it ran, for how long, at what load, and any RMA or repair records.
• Chain of custody and export compliance: advanced GPUs are export-controlled. Get a written statement of origin and ownership.
• Data sanitization: certificate of wipe for any drives that ship with the system.

On arrival: physical inspection

• Shipping damage, bent rails, cracked bezels.
• Dust load and corrosion on heatsinks and connectors.
• All fans and power supplies present and matching.
• For liquid-cooled systems: fittings, quick disconnects and any residue.

Don't reseat SXM GPUs yourself unless you are trained for it: the sockets are delicate and damage voids most warranties.

BMC and firmware

Log in to the baseboard management controller and export the System Event Log (SEL). Look for repeated thermal, power supply or memory events. Record BIOS, BMC, GPU VBIOS, NVSwitch and network card firmware versions, and plan updates to supported releases before production.

nvidia-smi checks

nvidia-smi                                   # all 8 GPUs visible, expected model and memory
nvidia-smi -q | grep -iE "serial|vbios|part number"
nvidia-smi -q -d ECC                          # volatile and aggregate ECC error counts
nvidia-smi -q -d ROW_REMAPPER                 # remapped rows, pending remaps, remapping failures
nvidia-smi -q -d PERFORMANCE                  # clock throttle reasons under load
nvidia-smi topo -m                            # NVLink topology between every GPU pair
nvidia-smi nvlink -s                          # link status and speed for each link
nvidia-smi nvlink -e                          # NVLink error counters

Reject or renegotiate if a GPU is missing, shows a row remapping failure, has uncorrectable ECC errors that recur, has links down, or throttles for power or thermal reasons at idle.

DCGM diagnostics

On HGX systems with NVSwitch, start the fabric manager first (systemctl status nvidia-fabricmanager). Then run NVIDIA's Data Center GPU Manager:

dcgmi discovery -l          # every GPU and NVSwitch found
dcgmi diag -r 3             # long diagnostic: memory, PCIe, NVLink, stress
dcgmi health -s a && dcgmi health -c

Level 3 runs for tens of minutes; run it at least twice. Any failure is a reason to stop and ask the seller.

NVLink and NCCL bandwidth

Measure GPU-to-GPU bandwidth with the open-source NCCL tests:

./build/all_reduce_perf -b 8 -e 8G -f 2 -g 8

Compare the reported bus bandwidth with a known-good system of the same model and the vendor's guidance. A result well below the baseline points to a degraded link, a missing NVSwitch or a firmware mismatch.

Burn-in

Run a sustained GPU stress test for 24 to 72 hours while logging temperatures, power and errors. Watch the kernel log for Xid errors:

journalctl -k | grep -i xid

Xid 48 (double-bit ECC), 63 and 64 (row remapping events and failures), 74 (NVLink error), 79 (GPU fell off the bus) and 94 or 95 (contained and uncontained ECC errors) all need an explanation before acceptance. Also test system memory, NVMe drives (nvme smart-log) and network cards at full speed.

Acceptance criteria to put in the contract

• All GPUs, NVSwitches and links present and healthy in DCGM level 3.
• No row remapping failures; uncorrectable ECC count zero after burn-in.
• NCCL bus bandwidth within an agreed percentage of the baseline.
• Burn-in completed with no Xid errors from the list above.
• A return window and a remedy (repair, replace or refund) if any check fails.

Questions