Acceptance checklist

Reported 8 GB and 10 GB

What to do when a card arrives, in order, so you can keep it with confidence or send it back inside the return window with evidence. The tests themselves are explained on Testing and burn-in. Why each step matters is on Buying.

Plan about an hour per card for steps 4 to 8, plus the time to install the unlock the first time.

1. Before it arrives

Save the listing, the seller's claims and the messages, on the platform you bought from. Keep all messages there; do not move to WhatsApp or email. Source: @darkmaniac7, 2026-10-26 (lab notes)
Note the return window and how returns are paid. Buy where the item can be returned, preferably for free. Source: @endsjeremy, 2026-10-14
Have a test host ready: Linux, the NVIDIA driver and CUDA toolkit, forced airflow over the card, and an 8-pin EPS power cable (see Cooling and power). Build the test tools in advance (see Testing and burn-in).
Do not solder the x16 cap mod until the card has passed. The mod only changes transfer speed; the tests do not need it. Source: repo github.com/dev-boz/cmp170hx-testsuite by dev-boz on GitHub, posted by @boz._., 2026-10-09.

2. Delivery

Film the unboxing in one continuous take, starting from the sealed parcel and its label. Source: @darkmaniac7, 2026-10-26 (lab notes) The test suite has a run sheet for filming a whole test session in one take. Source: repo github.com/dev-boz/cmp170hx-testsuite by dev-boz on GitHub, posted by @boz._., 2026-10-09.
Warning Photograph the shipping label and weigh the parcel. A member's multi-card order arrived as metal plates; the label declared a different item and about half the parcel's real weight. Source: @darkmaniac7, 2026-10-26 (lab notes)
Photograph any damage to the box. A card that arrived in a box with a hole and a bent bracket was no longer detected on PCIe; keep the packaging for the insurance or return claim. Source: @globalapathy, 2026-10-18
Source: @globalapathy, 2026-10-18

3. Look at the card

Bracket bent? Corrosion on the heatsink, rusted screws? Cards from farms with poor humidity control arrive oxidised; outside corrosion suggests the board may be corroding too. Source: @kellertreppe123, 2026-10-19
Source: @wakeupmbrc1, 2026-10-19
Dirt and missing thermal pads. Cards from Alifond arrived very dirty with pads missing on the back of the PCB and on side components. Source: @_syn_thetic_, 2026-10-25 Note it now; clean and repad after the card passes (Cooling and power).
The board shape: a real 170HX PCB is out of the PCIe CEM spec, with no key notch on the left of the gold fingers, so a different board in a 170HX shell is unlikely. Source: @dalzyyy, 2026-10-11
The power connector is present and undamaged. The card takes one 8-pin EPS input; see Cooling and power before using adapters.

4. First power-on and identity

Cap the power and run the fan at full speed for all tests. Source: repo github.com/dev-boz/cmp170hx-testsuite by dev-boz on GitHub, posted by @boz._., 2026-10-09. Test one card at a time; parallel tests make it hard to tell a bad card from a PSU, riser, board or cooling problem. Source: repo github.com/PixelML/club-170hx by PixelML on GitHub, posted by @c133, 2026-10-20.

lspci -nn | grep -i 10de                 # 10de:20c2 = 8 GB card, 10de:2082 = 10 GB card
lspci -nnv -d 10de: | grep -i subsystem
nvidia-smi -q | grep -i -E 'serial|vbios|board part|product name'
sudo nvidia-smi -i 0 -pl 180              # resets at every boot

Compare the PCI ID, subsystem, VBIOS and board part number with Which card do I have. A normal 8 GB card has board part number 900-11001-0108-000. Source: @zefrench6961, 2026-10-19 A 32 GB engineering-sample card reported subsystem 10de:1586. Source: @douya4305, 2026-10-13
Write down the serial number. It ties your test logs to this card.

5. Read the card's history

Before the unlock, read what the card has been through. Both tools only read.

sudo python3 blackbox.py live from gpcprobe dumps the InfoROM black box: boot count and power-on hours, power-cap history, and a log of Xid errors with timestamps. Source: repo github.com/matstatman/gpcprobe by matstatman on GitHub, posted by @matstatman, 2026-10-20. @globalapathy used @matstatman's tools the day they came out to avoid buying bad cards. Source: @globalapathy, 2026-10-20
sudo python3 gpc_probe.py shows which GPCs and memory partitions are disabled and which are fused defective. Source: repo github.com/matstatman/gpcprobe by matstatman on GitHub, posted by @matstatman, 2026-10-20.
10 GB cards: also read the RIR state with the read commands of 170hx-rir-tool (never the write command). Used cards whose RIR records were spent are expected on the market. Source: @dottoreargus, 2026-10-20

6. Unlock and check the result

Install the unlock as described on Driver unlock, then do a full power-off and power-on (a cold boot, not a reboot).

nvidia-smi --query-gpu=index,serial,memory.total,pcie.link.gen.current,pcie.link.width.current,power.limit --format=csv

8 GB card: memory.total 65536 MiB. 10 GB card: 40960 MiB. The test suite's identity check passes only on these values with no Xid since boot. Source: repo github.com/dev-boz/cmp170hx-testsuite by dev-boz on GitHub, posted by @boz._., 2026-10-09.
SM count: 74 on the 8 GB card with cmpunlocker v0.5. An older unlocker shows 70. Source: @matstatman, 2026-10-08
Source: @matstatman, 2026-10-08
PCIe: x4 is normal without the cap mod.
If a card stays at 32 GB, it is a single-rank or PG199-origin board; see Buying.

7. Run the tests

Run them in this order, one card at a time, as soon as possible after the cold boot. Details, commands and pass rules are on Testing and burn-in.

Memory: memtest_vulkan (5 minutes) right after a cold boot, then a CUDA test over the full unlocked VRAM.
Compute: gpu_burn -tc with 0 errors.
Tensor-core correctness: identical calculations compared across cards, or against a CPU reference. Cards exist that pass memory tests while their tensor cores return garbage. Source: @_____josh_____, 2026-10-08
Throughput: mmapeak or a PyTorch GEMM benchmark (BF16 or FP16 150 to 200 TFLOPS). Source: @jimtjames, 2026-10-11
A real workload that fills the upper VRAM and checks the output.
The kernel log: no Xid since boot.

A single passing quick memory test is not enough: a card with bad HBM passed a basic Vulkan test 9 times out of 10 and gave all-NaN inference. Source: @boz._., 2026-10-23

8. Decide

Warning Send the card back if:

memory size after the unlock is wrong (32 GB on an 8 GB card);
any memory test reports errors that are not a few fixed addresses (see Testing and burn-in#When a card fails);
gpu_burn reports errors, or the tensor results differ from other cards or the reference;
the card throws GSP or PMU Xids under load, or drops off the bus.

Some owners keep cards with a few bad fixed memory locations and work around them. Source: @analyticeth, 2026-10-08 That is a choice for after the return window, not a reason to skip the return.

If it fails:

Open the case on the platform early, inside the window, with the logs and the video. On Alibaba, a credit card chargeback recovered money where the platform dispute did not. Source: @djpandaman, 2026-10-04
Keep everything: the test suite saves raw logs per serial and date. Source: repo github.com/dev-boz/cmp170hx-testsuite by dev-boz on GitHub, posted by @boz._., 2026-10-09. The club-170hx QC guide lists what to keep: card label, driver and kernel, memory size, test pattern and duration, power limit, peak temperatures, PCIe width and generation, Xid and AER before and after. Do not publish serial numbers or full UUIDs. Source: repo github.com/PixelML/club-170hx by PixelML on GitHub, posted by @c133, 2026-10-20.

9. After the card passes

Clean, repad and repaste if needed, set up cooling, and make the power cap persistent (Cooling and power, Monitoring).
Do the x16 cap mod if you want it (Unlocks).
Keep the test logs. They are your baseline if the card degrades later.