Sell your GPUs
A provisional quote for every card, in minutes.
Tell us what you have. We email a one-time token, you run a read-only probe on the server, and a price for each card appears in your seller portal. Nothing leaves your site until you accept.
What happens next
- 01
We email a one-time token
Sent as soon as you submit, and tied to this request only.
- 02
Run one command
Paste it on each GPU server. It downloads ccprobe for Linux (x86_64 or arm64), checks its SHA-256, runs it and cleans up. It needs root to read kernel logs and NVML counters, and uses DCGM when it is installed.
$ curl -fsSL https://chip.bytebazaar.si/circuit/probe.sh | sudo sh -s -- --token 7QK4-M2ZP-9XH1 SHA-256 47d53fa6fd4ec42b… ok Found 8 GPUs · NVIDIA H100 PCIe 80GB · driver 550.90.07 Identity, ECC, row remapper, Xid history .... done Load test, 20 min · power, clocks, temps .... done dcgmi diag -r 2 ............................. pass Signing with session key (ed25519) .......... done Uploaded. Quote ready in your seller portal.
SHA-256 47d53fa6fd4ec42b9220a56fab14aef36ecfe0b9dd9754033959976d244ba077
- 03
See a price for every card
Plain-English findings per card (“2 correctable remapped rows: minor”) and a provisional price, valid for 7 days. Accept, counter or decline.
- 04
Sealed pickup, then the bench
We record tamper-evident seal numbers at pickup and again at receipt. The final price is set after our bench re-test.
What the probe reads
Hardware counters and kernel logs. Nothing else.
It never touches your files, models, datasets or network traffic. It reads, runs a 10–30 minute load test, uploads one signed report over HTTPS, and exits.
| What | Read from | Why it matters for your price |
|---|---|---|
| Identity: name, board part number, serial, UUID, VBIOS, InfoROM | nvidia-smi -q · NVML | Confirms the card is what it says. Catches relabelled cards and OEM-locked firmware. |
| Row remapper: correctable, uncorrectable, pending, failure | nvidia-smi -q -d ROW_REMAPPER | HBM health. A remapping failure means the card can’t be certified. |
| ECC counters: volatile and aggregate, SRAM and DRAM | nvidia-smi -q -d ECC | Shows wear in on-chip SRAM and in HBM. |
| Xid history: 48, 63, 64, 74, 79, 92, 94, 95, 119, 120 … | journalctl -k · dmesg | Past faults: fallen off the bus, NVLink errors, GSP errors, uncontained ECC. |
| PCIe link: generation and width, current vs max | nvidia-smi -q -d PCI · lspci -vv | A card at x8 instead of x16 points to a damaged slot or card. |
| NVLink status and error counters (SXM) | nvidia-smi nvlink -s -e | Baseboard and link health. |
| Power, clocks, temperatures, throttle reasons | NVML during a 10–30 min load | Condition of thermal paste and pads, and of power delivery. |
| NVIDIA diagnostics | dcgmi diag -r 2 | NVIDIA’s own pass or fail. |
| Compute and memory bandwidth vs reference | Built-in CUDA kernel · nvbandwidth | Catches slow stragglers that pass every other check. |
| Host: server model, CPU, driver and CUDA version | dmidecode · uname | Lets us quote the whole server if you want to sell it. |
Why the quote is provisional
You run the probe, so we check the probe.
A provisional quote lets you decide before anything ships. The final price comes from our own bench, on our own hardware.
- 01The report is signed with a key issued for your session and carries your token and a server nonce.
- 02If the same serial or UUID arrives from two sellers, both are flagged before anyone is quoted.
- 03Power, clocks and bandwidth must match the model’s baseline. An emulated or relabelled card can’t fake them.
- 04If a card tests lower on our bench than its probe suggested, we re-quote that card before paying. Rejected cards go back to you or to an EPR-registered recycler.