TL;DR: GeForce RTX cards run AI training math on their tensor cores at half the rate NVIDIA's workstation cards get from the same chip. gpu-unlock, a free open-source Linux module, lifts that cap at every boot. Games run the same.
Get it from one place
The only official source is github.com/Zagitta/gpu-unlock. The developer says websites, installers and coins that use the name are not part of the project. Preporato isn't affiliated with it and hosts no files.
What it does
NVIDIA's own whitepaper lists the same GA102 chip at two rates:
Same chip, two rates
FP16 tensor math with FP32 accumulation (NVIDIA GA102 whitepaper), with each card's launch price
- RTX A6000 ($4,650)154.8 TFLOPS
- RTX 3090 ($1,499)71.0 TFLOPS
The limit lives in two settings inside the chip, called speed selectors. gpu-unlock sets them before NVIDIA's driver loads:
Why the limit exists, in the developer's words: "Market segmentation, they want businesses to purchase the professional cards without these limitations at significantly higher prices."
What you get
BF16 matrix multiply, TFLOPS
The developer's numbers: dense matrix multiply with FP32 accumulation, stock power limits
- RTX 3060 12 GB27 51+87%
- RTX 307040 72+81%
- RTX 3080 10 GB65 113+73%
- RTX 309075 125+65%
- RTX 3090 Ti (450 W)88 157+80%
- RTX 4070 Ti81 150+84%
- RTX 4080104 193+87%
- RTX 4090165 329+100%
The RTX 4090 runs out of power before it runs out of tensor cores, so large jobs at stock power gain much less than its peak rate.
In real work, from the compatibility reports users have filed:
| Workload | Typical speedup |
|---|---|
| LLM prompt processing (prefill), FP16/BF16 or AWQ/GPTQ | +25 to 55% |
| LoRA and full fine-tuning in BF16 | +25 to 50% |
| Image generation in FP16/BF16 (SDXL, SD 1.5, ComfyUI) | +30 to 55% |
| Video generation in BF16 (Wan 2.2) | +15 to 20% |
| llama.cpp and Ollama with quantized GGUF models | 0 to 15% |
| Token generation (decode), INT8/FP8/INT4 models | No change |
| Games and FP32 CUDA | No change |
The results are bit-identical before and after: only the speed changes.
Which cards work
3050 · 3060 · 3060 Ti · 3070 · 3070 Ti · 3080 · 3080 Ti · 3090 · 3090 Ti
Plus the 3070 and 3080 Laptop GPUs. 100+ user reports.
4060 · 4060 Ti · 4070 · 4070 SUPER · 4070 Ti · 4070 Ti SUPER · 4080 · 4080 SUPER · 4090 · 4090 D
Plus the 4050 and 4060 Laptop GPUs.
2060 · 2060 SUPER · 2070 · 2070 SUPER · 2080 · 2080 SUPER · 2080 Ti · Titan RTX
The developer calls it unlikely to happen.
Laptops
A laptop stays unlocked only while its dedicated GPU never fully powers off, and many laptops power it down when idle. The README suggests options nvidia NVreg_DynamicPowerManagement=0x00 in a file in /etc/modprobe.d/, which helps on some models.
Before you start
You need Linux with NVIDIA's driver: the Debian and Ubuntu family (Mint too), Arch (CachyOS too), or Fedora and RHEL with DNF. On Fedora Atomic (Silverblue, Kinoite, Bazzite) make install doesn't work; the open pull request #79 describes the manual steps.
Secure Boot
With Secure Boot on, Linux refuses an unsigned kernel module. Either enroll the key DKMS signs it with (mokutil --import, then a reboot through the MOK manager, as the open pull request #174 describes), or turn Secure Boot off in your PC's firmware settings.
Install on Linux
Tick each step as you go
0/6 completed1. Download and install it
git clone https://github.com/Zagitta/gpu-unlock
cd gpu-unlock
make install
make install detects your distribution, installs what the build needs and asks for your sudo password. The build downloads NVIDIA driver files to generate the firmware the unlock uses, so it needs an internet connection.
2. Switch it on
The module does nothing until your boot line asks for it. Add gpu_unlock.targets=all and keep the options that are already there.
Open /etc/default/grub and add the flag inside GRUB_CMDLINE_LINUX_DEFAULT:
GRUB_CMDLINE_LINUX_DEFAULT="quiet splash gpu_unlock.targets=all"
Then update GRUB:
sudo update-grub
To unlock only some cards, list their PCI addresses instead of all, as lspci -D prints them:
gpu_unlock.targets=0000:21:00.0,0000:22:00.0
3. Restart
The unlock takes about a second. The screen goes black for a moment while the card resets, until NVIDIA's driver takes over.
Check that it worked
journalctl -k -b | grep gpu_unlock
It says which cards were unlocked, and why a card was skipped or refused. lspci -nnk -s 0000:01:00.0, with your card's address, shows that NVIDIA's driver is bound to the card again.
Undo it
In GRUB, press e on the boot entry, delete gpu_unlock.targets=all, and boot with Ctrl+X or F10.
A full power off always returns the card to stock. After suspend to RAM the module unlocks the cards again; hibernation isn't covered.
On Windows
No official Windows version
The developer: "I don't run windows and don't plan to maintain it", and "There are windows installer's floating around but I would recommend being very careful with those."
Two community projects unlock the card in a boot step before Windows starts:
- nv-uefi-unlocker runs from a FAT32 USB stick you pick in the one-time boot menu. It needs UEFI boot and Secure Boot off (or a self-signed setup), lists RTX 20, 30 and 40 cards as compatible, and the cap comes back at power off. Its steps are in the English README.
- Windows-GPU-UNLOCK is the port the developer links from the FAQ. It describes itself as an RTX 3090 UEFI port.
We haven't tested either. Turning Secure Boot off lowers your PC's protection against malware that loads at boot, and both projects started in October 2026, so read their issues first and download only from their GitHub pages.
Is it safe for the card?
The developer calls harm "very unlikely": the card may run slightly warmer because it does more compute, and NVIDIA's usual throttling and power management stay in place. Nothing is written to the card, so a full power off returns it to stock. RTX 20 and 40 support is still experimental; if a card fails to unlock, the kernel log says so.
Put the speed to work
Real AI projects on a cloud GPU, step by step: fine-tune Llama 3 8B with LoRA, fine-tune Stable Diffusion, train a small language model, and see what FP16 and BF16 do to speed.
Fine-Tune an LLM with LoRA and QLoRA (Jupyter)
Fine-tune Meta Llama 3 8B on a custom instruction dataset using LoRA and QLoRA. Learn parameter-efficient fine-tuning from data preparation through evaluation — the #1 most demanded AI skill.
Fine-Tune Stable Diffusion with LoRA: Custom Text-to-Image
Load Stable Diffusion, attach LoRA adapters to the U-Net's attention layers, run a tiny overfit training loop, and generate with the adapted weights to prove that a few million trainable parameters actually move pixels.
Train a Small Language Model from Scratch
Train a real GPT-style language model from zero on TinyStories: tokenize, wire up the optimizer and LR schedule, run the training loop with validation perplexity, and generate coherent text from your own weights. End-to-end pretraining in minutes on one GPU.
Batch Size & Precision Sweep: Finding Your Sweet Spot
Sweep batch sizes and numerical precisions (fp32, fp16, bf16) on a real model to find the throughput/VRAM knee, then ship a production recommendation with SKU-aware precision picks and an accuracy gate.
Frequently asked questions
Sources
- gpu-unlock README, read on 2026-11-05: github.com/Zagitta/gpu-unlock
- NVIDIA Ampere GA102 GPU Architecture whitepaper, Tables 3 and 9: nvidia.com
- gpu-unlock pull requests #79 (Fedora Atomic) and #174 (Secure Boot)
- nv-uefi-unlocker README (English): github.com/physicsdolphin/nv-uefi-unlocker
- Notebookcheck, on gaming: Nvidia RTX 3090 performance nearly doubles with free GPU unlock, but there's a catch