Files

T

qbuissondebon ac53b83f70 Name card folders by GPU UUID, record serial in serial.txt

Folder identity switches from the physical serial (which neither
nvflash nor nvidia-smi exposes on the 170HX) to the auto-read GPU
UUID from nvidia-smi. The physical serial probe result (or N/A) is
saved to <folder>/serial.txt and report.json. No manual serial
prompt anymore; --serial only overrides the recorded serial.
Unidentified runs fall back to unidentified/<bdf>/. AGENTS.md
invariant and README updated to match.

2026-10-15 00:17:13 +02:00

3.6 KiB

Raw Blame History

Project instructions

Purpose

This repository checks one NVIDIA CMP 170HX at a time on a fresh Ubuntu 26.04 x86-64 host. The normal sequence is: prepare driver and tools, check PCIe lane width, install cmpunlocker for stock memory, cold boot, confirm unlocked memory, then check the final PCIe link. The capacitor width check passes at x16, even if the link is Gen1. A completed post-unlock card check passes only at Gen2 x16.

Source and outputs

check_cards.py owns the CLI, per-card flow, report scoring, and terminal result.
cxlib.py owns host setup, the cmpunlocker harness clone/update and opt-in installer, card identity, subprocess logging, PCIe inspection, and optional VBIOS backup.
README.md is the operator guide. Keep its flags, output paths, reboot steps, and exit codes aligned with the code.
harness/, cards/, and unidentified/ contain generated source, logs, reports, or firmware backups. Do not commit their contents.
The bundled nvflash_5.867_linux/ contains the optional backup tool. Backups belong under extras/, not at repository root.

Invariants

A card folder is named from the automatically read GPU UUID (nvidia-smi). The physical serial (nvflash probe, else N/A) is recorded in <folder>/serial.txt and report.json. Never prompt for manual serial entry. When even the UUID is unavailable, continue the run under unidentified/<bdf>/ and report it.
Match every card-sensitive operation to the detected PCI address. A host may have another NVIDIA GPU even though only one CMP 170HX is being checked.
Measure capacitor-related lane width before unlocking. The width result is independent of cmpunlocker: Gen1 x16 passes the width check, while x4 fails it. Do not claim the capacitor is physically verified; the width is an observed link result on the current host.
Keep the final PCIe acceptance check after unlock. Preserve a red [ERROR] message and failing exit for any final link other than Gen2 x16. Gen1 x16 passes the earlier width check but fails the final card check.
Preserve cold-boot state by card identity. GPU UUID may be used internally for this purpose, but not as a folder name.
The cmpunlocker harness is cloned from upstream main into harness/cmpunlocker/, and git pull --ff-only runs before every install (a failed pull warns and continues with the existing checkout). Invoke install.sh only with flags the checked-out version actually accepts — currently no flags; never pass --no-passthrough to older checkouts without verifying them first.
VBIOS backup is optional and read-only. Select the detected PCI address for nvflash and sysfs ROM operations. Never add a flashing step to the normal workflow.

Working in this repository

The repository has a single remote, origin (git@git.ssh.qbuissondebon.eu.org:qbuissondebon/cmp170hx_check.git, web: https://git.qbuissondebon.eu.org/qbuissondebon/cmp170hx_check), default branch main. Local status, diff, and commits are allowed at any time. Push to origin only when the user explicitly asks for it. Never force-push, never rewrite pushed history, and never add another remote without an explicit request.
Do not run host setup, unlock, VBIOS, or PCIe mutation commands on a development machine. Those actions belong on the intended Ubuntu host when the user requests them.
Keep reports and long-running installation logs visible and useful to an operator. Save full output on disk; show installation progress in the terminal.
If asked to verify without a card, python3 check_cards.py --demo is the synthetic path. It does not establish hardware correctness.