GitHub radar
Unlock faster AI compute on your Nvidia RTX 30-series GPU
This is a Linux kernel module that removes a built-in speed limit on Nvidia RTX 30-series cards, making local AI models run noticeably faster.
A tiny program loads into Linux before your graphics driver starts and lifts a factory compute speed limit.
When it helps
Useful if you have a desktop RTX 3060–3090 Ti on Linux and run AI models locally; it won't work on Windows or turn your card into a pro-grade one.
Pros
- Doubles FP16/BF16 compute speed at no extra cost
- No new GPU or driver needed
- Keeps working automatically after every reboot
- Already near 11,000 Github stars within a week
Cons
- Works only on Linux, not Windows
- Changes core kernel and bootloader settings
- GPU may run slightly hotter under full load
How to set it up — step by step
- 1Check your GPU: you need a desktop RTX 3060–3090 Ti on Linux, not Windows.
- 2Open your AI agent (Claude Code or Codex) in an empty folder on a spare Linux machine.
- 3Send the agent the text from the block below to learn the risks before installing anything.
- 4To install yourself: clone the repository and run the build command make install.
- 5Add gpu_unlock.targets=all to your bootloader settings (GRUB or similar) and reboot.
- 6Check the result with the journalctl command from the README and compare model speed.
Text for your agent
Copy this and send it to your agent — Claude Code, Codex, any of them:
Read the README at https://github.com/Zagitta/gpu-unlock and explain in plain terms what this means for my GPU: what it actually changes, what risks the author himself names, and whether this unlock is even worth considering for me — do not install or build anything.
You need a desktop Nvidia RTX 30-series card on a GA102, GA103, GA104 or GA106 chip (RTX 3060 through 3090 Ti) and Linux — this is a kernel and bootloader modification, not an ordinary app; the author does not officially support Windows.
Open on GitHub▌ More finds