Zagitta/gpu-unlock

The problem

Nvidia's GA102/GA103/GA104/GA106 dies ship with full-rate FP/BF16 tensor issue, but GeForce cards have it fused down to half rate. The only way to get it back has been a UEFI application or a patched driver — both fragile: the UEFI path needs manual firmware work, and the driver patch has to be redone for every new nvidia release. This project is a kernel module that performs the unlock in the initramfs, before the unmodified nvidia driver binds the card, so nothing needs repatching when drivers update.

What's different

It combines the CMP unlocker's V67/PLM technique with the pre-driver timing of the UEFI approach, and packages it as a normal Linux kernel module with distro package targets. That means the unlock survives nvidia driver updates by design, and it participates in the standard module lifecycle (suspend/resume, multi-card, vfio passthrough). The trade-off versus a patched driver is scope: it only helps on the listed consumer chips, only on Linux, and only when the module loads before nvidia. Versus the UEFI route it removes the manual firmware flash but adds a kernel module in your initramfs, which is a real trust and stability surface.

Who it's for

Linux users on RTX 30-series cards (reported: 3090 Ti, 3090, 3080, 3080 Ti, 3070, 3060 12 GB) who run matrix-heavy work — LLM prefill, diffusion, FlashAttention, GEMM — and who build their own kernel modules. It is not for LLM decode or INT8 workloads, which stay bandwidth-bound and see little gain; not for Windows users (the project is Linux-only); and not for anyone who wants a supported, warranty-safe GPU. Cards not on the compatibility list, and GA103, have no reports yet.

Cost of adoption

Linux only, with build tooling for Arch, Debian/Ubuntu and Fedora/RHEL; other distros use the explicit targets. Install is make install followed by a kernel command-line opt-in (gpu_unlock.targets=all or specific PCI addresses), plus bootloader edits. The module embeds in your initramfs, so the display goes blank briefly at boot and resume takes about a second longer. The license is not declared, which is a real question for anyone packaging or redistributing it. There is no long-term track record: 6 days old with hundreds of stars per day.

Verdict

This is a brand-new project riding a sharp spike — 2,766 stars and #17 on the trending board in its first week, with an active issue tracker of user benchmark reports — not an established tool. The modest gain numbers come from users' own benchmarks, not a controlled suite, and the +25% floor shows power limits can eat much of the benefit. It is genuinely useful for a narrow audience with the right chips and the willingness to run an unsigned-ish kernel module in initramfs, but the missing license and six-day age mean treat it as experimental: measure before you rely on it. If you need a supported, long-lived setup, stay on a patched driver or accept the stock half rate.