A used GPU server comes with a history. Its GPUs have corrected memory errors, remapped faulty rows and logged driver errors, and its warranty terms were set when someone else bought it. On Ampere and later NVIDIA GPUs, much of that history can be read before you buy: nvidia-smi keeps aggregate ECC counts as a lifetime counter and does not support clearing them, and remapped rows are recorded in each GPU's InfoROM, a small persistent store on the board [1, 2].
This explainer names the records to ask for and what NVIDIA, two server makers, NIST and the Bureau of Industry and Security (BIS) say about each. Prices are a separate question, covered in what sets the price of a used H100 server, and so is the bill of materials, set out in what an HGX server price includes.
What to check on a used GPU server
The checks fall into seven groups: identity and configuration, memory health counters, NVIDIA's return criterion, error logs, diagnostics, warranty, and drives with export screening. One published reuse standard, R2v3 from Sustainable Electronics Recycling International, frames the first question for any quote. Its test and repair appendix recognizes organizations that test, repair or refurbish equipment and "accurately communicate the level of functionality, cosmetic condition, and data sanitization status", and it says a power-on check is not a final indication of a device's level of functionality [30, 31]. So a quote should say what was tested, what condition the unit is in and what happened to the data on its drives.
Serial numbers and configuration as received
nvidia-smi reports a serial number that matches the label printed on each board, and a GPU UUID that matches no label at all [1]. On a board that carries several GPUs, the serial "might be shared between multiple GPUs on the same board" [1], so a unit report should list both for every GPU. The same output gives the VBIOS version, part numbers, the chassis serial and the InfoROM image version, and the query field inforom.checksum_validation reads the InfoROM checksum as valid, invalid or N/A [1].
Counts follow from the baseboard. NVIDIA's Fabric Manager guide gives an HGX H100 8-GPU baseboard eight GPUs and four NVSwitches, an HGX A100 eight GPUs and six NVSwitches, and HGX B200 and B300 eight GPUs and two NVSwitch chips [4]. What a used DGX includes beyond an HGX baseboard is set out in DGX, HGX and MGX explained, and the generations side by side in HGX H200, B200 and B300 compared.
Read PCIe links with the GPU busy, since nvidia-smi's current link generation and width "may be reduced when the GPU is not in use" [1]. Record the management controller's firmware too. CVE-2023-25528, published in September 2023, was a flaw in a web server plugin of the DGX H100 baseboard management controller (BMC) through which an unauthenticated attacker may cause a stack overflow, and the National Vulnerability Database scored it 9.8 [33].
How to check used GPU health: ECC counts and remapped rows
From Ampere on, every DRAM bank has spare rows, and a faulty row is remapped to one of them. A remap needs a GPU reset to take effect and persists for the life of the GPU, with up to 512 remaps for the frame buffer [17]. nvidia-smi reports active and inactive remaps for correctable and uncorrectable errors, a Pending field and a Remapping Failure Occurred field [1]. NVIDIA notes that the remapped-row count is the number of entries recorded in the InfoROM [2].
A bank remap availability histogram sorts each memory bank into Maximum, High, Partial, Low or None [1]. NVIDIA's page shows examples, such as 640 banks at Maximum, and notes that the per-bank availability column in them is not available through its APIs or nvidia-smi [2]. Treat those figures as examples from one page [2].
A pending flag means a remap takes effect at the next GPU reset [19]. Ask the seller to reset the GPUs and read the counters again before the report is issued.
One fleet study gives the memory counters extra weight on Hopper units. The Delta study followed 448 A100s for 895 days and 608 H100s for 146 days, and found the H100s had a 3.2 times lower mean time between uncorrectable memory errors per GPU [8]. The same study found the H100s more resilient in other critical hardware, with only three GPU system processor (GSP) error cases and no NVLink errors in their 146 days [8]. Its H100s sit in GH200 Superchips, which the authors distinguish from discrete H100 GPUs [8].
NVIDIA's return criterion for remapped rows
NVIDIA's return (RMA) criterion for row remapping is a set row-remapping failure flag, validated by NVIDIA's field diagnostic [3]. The flag is set by an uncorrectable-error remap attempt in a bank that already has eight such remaps or on a row already remapped, or once 512 uncorrectable-error remaps have occurred in total [3]. On Blackwell GPUs, a third uncorrectable remap attempt in one bank invokes channel repair in its high-bandwidth memory when a spare channel is available [3]. NVIDIA calls the field diagnostic its authoritative tool for GPU health and says it is usually required before a return can start [22].
No NVIDIA page reviewed sets a lower count of remapped rows or ECC errors that makes a GPU unfit. A buyer who wants a stricter limit writes it into the purchase terms.
Error history lives in logs
Xid messages are the NVIDIA driver's error reports, printed to the operating system's kernel log [5]. NVIDIA ties Xids 48, 63, 64, 92, 94 and 95 to GPU memory errors [22], and Xid 79, a GPU that has fallen off the bus, often to hardware failures on the PCIe link [21]. When a GPU issue is reported, NVIDIA asks for nvidia-bug-report.sh output, Fabric Manager logs on HGX systems and DCGM diagnostic logs [22]. Because Xids are written to the operating system's logs [20], ask for logs captured before the operating system was reinstalled, and for the management controller's own event log [6].
Host parts matter too. In a 2024 study of two research clusters of A100 servers, GPUs caused 28.2% of the faulty "lemon nodes" behind repeated job failures, DIMMs 20.5%, PCIe 15.4% and NICs 7.7% [7]. The study found those nodes with signals such as repair tickets and the number of times a node was taken out of service, and its authors state that historic data is necessary to find defective nodes [7].
Diagnostics to run or require
NVIDIA's Data Center GPU Manager (DCGM) diagnostics check PCIe, NVLink, memory, compute, power and thermal stress and memory bandwidth before a workload runs [24]. On 8-GPU Hopper systems, NVIDIA measured level 3 at under 35 minutes and level 4 at under 2.25 hours [24]. NVIDIA states that these diagnostics are not designed to replace its field diagnosis tools or to facilitate an RMA [24]. Ask which level was run, and for its logs.
Warranty: who stands behind the unit
Supermicro's published warranty applies only to the original purchaser and excludes components supplied by third parties. For GPU systems it lists three years of labor and one year of parts and advance cross-ship, unless an invoice sets special terms [9]. Dell's system warranties may be transferred when the current owner records the transfer with Dell, at Dell's sole discretion, and warranty support applies only in the country where Dell originally sold the system [10]. NVIDIA support entitlements are non-transferable and may be terminated when a product changes hands [11], and any user can check an NVIDIA product's warranty on NVIDIA's Warranty Check form [25]. Ask whose warranty applies, who provides it and when it ends.
Drives, management controllers and export checks
NIST SP 800-88 Rev. 2 names resale among the points where control of storage media changes, prefers purge to clear where possible and gives a sample certificate that records each drive's serial number, method, tools and verification [12]. Ask for one record per drive. On arrival, reset the BMC's users and credentials and update its firmware, in line with CISA and NSA guidance on hardening BMCs [32].
Export guidance draws no line between new and used units [13]. BIS guidance of 13 May 2025 names servers by export classification and lists red flags such as an unknown delivery or installation address [13]. A seller that screens its sales asks about the end user, the end use and the installation site [13, 34]. This is general information, not legal advice.
Rillor grades and tests surplus, refurbished and used GPU servers before sale, and each unit comes with its own report: serial numbers for the system and each GPU, the configuration as received, firmware versions, burn-in and diagnostic results, physical condition and the condition grade.
Read the full report
The full report adds the Xid codes tied to memory errors, the DCGM run levels, the Supermicro, Dell and NVIDIA warranty terms, the NIST certificate fields, the BIS and Know Your Customer red flags, a buyer's checklist by stage and all 35 sources.