GPU Unlock is a free, open-source Linux kernel module licensed GPL-2.0-only. Its developer says it can restore higher FP16, BF16 and TF32 tensor-math throughput on specifically supported GeForce RTX 20-, 30- and 40-series cards. It runs during boot before the standard NVIDIA driver loads; it does not flash the VBIOS or replace the NVIDIA driver. Its official README reports faster results in some AI workloads, but those figures are user reports—not a promise that every application will run twice as fast.

Important: installing a kernel module and changing boot parameters are system-level operations. Check your exact card and Linux boot setup, review the official repository, and back up important files and boot settings first. This guide explains the project’s published steps; it does not independently test or recommend installing the module.

Which GeForce cards are listed as compatible?

The upstream compatibility list distinguishes stable from experimental support. “Stable” does not guarantee the same result on every board, laptop, BIOS or Linux configuration.

GPU generation Cards listed by the project Project status
RTX 30 / Ampere RTX 3050, RTX 3060, RTX 3060 Ti, RTX 3070, RTX 3070 Ti, RTX 3080, RTX 3080 Ti, RTX 3090 and RTX 3090 Ti; RTX 3070 and RTX 3080 Laptop GPUs Stable; upstream README cites 100+ user reports
RTX 40 / Ada Lovelace RTX 4060, RTX 4060 Ti, RTX 4070, RTX 4070 SUPER, RTX 4070 Ti, RTX 4070 Ti SUPER, RTX 4080, RTX 4080 SUPER, RTX 4090 and RTX 4090 D; RTX 4050 and RTX 4060 Laptop GPUs Experimental
RTX 20 / Turing RTX 2060, RTX 2060 SUPER, RTX 2070, RTX 2070 SUPER, RTX 2080, RTX 2080 SUPER, RTX 2080 Ti and Titan RTX Experimental
RTX 50 / Blackwell No models listed Not supported

There is an architecture-specific caveat for RTX 20 cards: Turing GPUs do not have BF16 tensor hardware. Tom’s Hardware notes that Turing unlock results therefore concern FP16, not BF16. On laptops, the project says the unlock can reset if the dedicated GPU powers fully off while idle, so results may be mixed. Check the upstream compatibility list for the latest status and exact device IDs before proceeding.

What the module changes—and what it does not

The project says the module applies its change early in the boot process, before the stock NVIDIA driver binds to the GPU. This is not a VBIOS flash or a patched driver, and the developer says the stock driver can continue to load normally afterward. The project is Linux-only and is a community-developed tool, not an NVIDIA utility. The compatibility table above is the developer’s own list.

Linux installation steps

The commands below follow the project README and the supplied setup guide. The README says make install detects common Arch, Debian/Ubuntu and Fedora/RHEL-family systems. Do not replace your existing kernel options: add the new parameter to them.

  1. Clone the official repository and install:
    git clone https://github.com/Zagitta/gpu-unlock
    cd gpu-unlock
    make install
  2. Add the boot parameter. Append gpu_unlock.targets=all to the kernel command line. Keep the options already there.

    For Debian, Ubuntu or Mint using GRUB, edit /etc/default/grub and add the parameter inside the existing GRUB_CMDLINE_LINUX_DEFAULT value, then run:

    sudo update-grub

    For Arch or CachyOS using GRUB, make the same edit, then run:

    sudo grub-mkconfig -o /boot/grub/grub.cfg

    For Fedora or RHEL, the project documents:

    sudo grubby --update-kernel=ALL --args="gpu_unlock.targets=all"

    With systemd-boot, add the parameter to the options line in the boot entry. The upstream README says bootctl status can show the entry location. For a Unified Kernel Image (UKI) or generated entries, add it to /etc/kernel/cmdline and regenerate the boot images using your distribution’s process.

  3. Restart so the module can run during boot, before the NVIDIA driver loads.

These are boot-level changes, and a wrong kernel-command-line edit can affect startup. If your distribution or bootloader is not listed, consult its documentation and the project’s current setup instructions rather than guessing.

How to check the result and undo it

After restarting, the guide suggests checking the current boot’s kernel log and the NVIDIA device’s driver binding. Replace [ID] in the second command with the correct PCI address for your GPU.

journalctl -k -b | grep gpu_unlock
lspci -nnk -s 0000:[ID]:00.0

To stop applying the unlock, remove gpu_unlock.targets= from the kernel command line and reboot. For a one-time boot, remove the parameter in the bootloader’s edit screen. The upstream README says removing the installed package also removes the module and its initramfs hooks. Follow your distribution’s package-removal procedure; the guide does not provide a universal uninstall command.

What performance does the project report?

The project README labels these workload gains as typical results from user compatibility reports. They are not controlled, universal benchmarks. The supplied guide lists:

Workload Reported result Workload Reported result
Image generation: SDXL, Stable Diffusion 1.5 and ComfyUI 30–55% faster Video generation using BF16, including Wan 2.2 15–20% faster
LLM prefill in vLLM or PyTorch 25–55% faster llama.cpp or Ollama with GGUF 0–15% faster
LoRA or fine-tuning in BF16 25–50% faster Token generation with INT8, FP8 or INT4 No change reported
FlashAttention or SDPA 50–60% faster Gaming and FP32 CUDA No change reported

The current README also reports a 25–30% lower time-to-first-token in the configurations represented by its user reports. It says outputs are bit-identical before and after the change, and that only speed changes; treat that as the developer’s statement, not an independent guarantee for every software stack.

There is an important difference between doubling a GPU’s peak tensor rate and doubling an end-to-end app’s speed. Tom’s Hardware’s independent coverage distinguishes peak throughput from application results and reports that performance varies by card, workload, power limit and project or fork tested. It also notes that memory-bound LLM token generation and some quantized paths may show little or no improvement. The GPU Unlock README likewise warns that power and thermal headroom matter; it specifically says the RTX 4090 can be power-limited.

Safety notes and Windows alternatives

Use the official Zagitta repository as the source for the Linux project. Its README says the GitHub repository is the official source and that other websites, coins or installers are not affiliated with the developer; it warns users to be very careful with Windows installers circulating elsewhere. A kernel module runs with high system privileges and this procedure changes boot-time behavior, so review what you install and keep a recovery plan. The project says the GPU may run slightly warmer under load, while normal throttling and power management remain active.

The guide names two separate community Windows/UEFI projects: Windows GPU UNLOCK and nv-uefi-unlocker. They are not maintained by the original Linux project’s developer and have their own hardware compatibility and recovery instructions. They are not covered by the Linux commands in this article.

Frequently asked questions

Does GPU Unlock double every AI application’s speed? No. The project claims up to double peak tensor throughput for supported math modes, but reports vary by workload. Its own compatibility reports list no change for some token-generation, quantized, gaming and FP32 workloads.

Does it support RTX 50 cards? No. The upstream compatibility table lists RTX 50 / Blackwell as unsupported.

Is there a Windows installer from the original developer? The original project describes itself as Linux-only and says it has no associated installer. The community Windows/UEFI projects are separate; use their own compatibility and recovery documents and do not confuse them with the Zagitta Linux module.

Sources: GPU Unlock official README and compatibility reports; Tom’s Hardware’s independent report; Windows GPU UNLOCK; nv-uefi-unlocker.