Audited 2026-10-25 · Sources: GitHub repo metadata + README + issue reports, watcher security scan, HN/X activity reported in our digest. Tabletop audit — nothing installed or executed, and nothing in this article is a reason to put it on a machine you care about.
The card
| Repo | Zagitta/gpu-unlock |
| Stars | ~7.0k (GitHub, audit date) — the digest caught it at 2,367 and it nearly tripled inside the watch window. Top-15 Hacker News thread, a ~450k-view X post, and a community RTX 4090 port (gpu-unlock-ada) already exists. |
| License | NONE — no license file. Default copyright applies: no explicit right to modify, redistribute, or build derivatives. The 4090 fork is already operating in that gray zone. |
| Provenance | Single-author repo (Zagitta), created 2026-10-18. Combines the published CMP-unlocker V67/PLM technique with pre-driver timing from the UEFI approach — not a patched NVIDIA driver and not a UEFI app. Per-card results are attributed to numbered issue reports. |
| What it is | A Linux kernel module that loads from initramfs before the NVIDIA driver binds, writes the card's speed selectors (SS0/SS1), and performs a function-level reset — unthrottling the FP/BF16 issue rate on GA102/103/104/106 (RTX 3090-class) silicon toward professional-card rates. |
| Install surface | Maximum: a kernel module in your boot chain plus a kernel command-line flag. make install detects your distro, builds a signed-format package, and asks for sudo. There is no larger trust grant on this machine. |
What problem it solves
NVIDIA rate-limits FP16/BF16 FMA issue on GeForce cards — a market-segmentation cap, not a hardware limit. The author's own 3090 Ti numbers: dense GEMM BF16 reportedly up to ~157 TFLOPS from ~87, TF32 roughly doubled, with user reports putting BF16 gains at up to +80% on a 3070. Where it helps: LLM prefill (reported +14–36%), diffusion, FlashAttention. Where it doesn't: LLM decode is memory-bandwidth-bound and stays flat — chatbot snappiness is unaffected.
Our gate findings
| Gate | Finding |
|---|---|
| Upstream identity | Thin-to-moderate — one author, a week-old repo, no company. The corroboration is the issue tracker: multiple users reporting measured results per card. |
| License | FAIL — none. No license means no granted rights. Whatever the code does, the legal posture alone fails the wall. |
| Claims discipline | Better than expected. The results table links every figure to its issue report, includes a 3080 Ti OEM variant where the unlock completes but BF16 stays at half rate — a published partial failure — and warns that power caps skew results (the lowest report, +25%, came from a 225 W-limited card). |
| Build hygiene | Genuinely good: all inputs are pinned and fetched over HTTPS with SHA-256 verification, and the dist tarballs contain no NVIDIA archives or generated firmware. This is the part that usually fails first; here it's the strongest gate. |
| Blast radius | Everything. A pre-driver kernel module writing GPU speed selectors runs at the highest privilege the machine offers, and a mistake is a bricked boot or a cooked card — not a bug report. An unclean shutdown path or a future NVIDIA driver change is yours to debug. |
| Security scan | WARN, zero failures — the only warnings were two scanner tools missing on the runner. But a clean scan is not clearance: the risk here is architectural (ring-0, pre-boot, hardware registers), not malware-shaped. |
Verdict: NO-GO for machines you depend on. Honest README, clean scan, real user reports — and it still fails the wall: no license, no vendor posture, and a trust grant at ring 0. If anyone experiments, it's a sacrificial bench box with a recovery plan and the owner's explicit OK — which is the answer we gave when it hit our own queue.
Best for
Homelab ML tinkerers with a spare Ampere card and a reflash plan — and everyone else as a case study in "the documentation is honest and the answer is still no."