GitHub radar
Double FP16/BF16 Speed on RTX 4090 With a Linux Kernel Module
This is a Linux kernel module that unlocks a hidden fast mode on the RTX 4090, nearly doubling FP16/BF16 tensor throughput — useful if you run AI models locally.
A Linux kernel module (a low-level system extension) that turns on a hidden performance mode on the RTX 4090 before the NVIDIA driver loads.
When it helps
Useful if you have an RTX 4090 (or, experimentally, an RTX 4070 Ti SUPER) on Linux and run AI models locally; skip it on other cards, on Windows, or if you don't run models yourself.
Pros
- Nearly 2x FP16/BF16 throughput, confirmed independently by several RTX 4090 owners
- Speeds up the GPU's compute units themselves, not just one app or model
- Open source, builds on your own machine straight from source
Cons
- Modifies GPU firmware and the Linux kernel — risks your warranty and boot stability
- Showed no gain on a quantized MoE model in the authors' own test
- Support for cards other than the RTX 4090 is experimental, tested on 1-2 units
How to set it up — step by step
- 1Check your GPU: run `lspci -D` and confirm it's RTX 4090 (`10de:2684`) or, experimentally, RTX 4070 Ti SUPER.
- 2Install headers for your exact Linux kernel version and a working DKMS toolchain — the module won't build without them.
- 3Clone https://github.com/JustAResearcher/gpu-unlock-ada and run `make install`.
- 4Power the machine fully off and start it again for the first run — a normal reboot may not be enough.
- 5Send the agent the text from the block below to learn the pros, cons, and risks for your card.
- 6Test the gain on your own workload — quantized models showed no improvement in the authors' tests.
Text for your agent
Copy this and send it to your agent — Claude Code, Codex, any of them:
Read the README of the JustAResearcher/gpu-unlock-ada repository (https://github.com/JustAResearcher/gpu-unlock-ada) and tell me the pros, cons and risks of this kernel modification — whether it even fits my card, if I have an RTX 4090 or RTX 4070 Ti SUPER on Linux.
This is not ordinary software. It is a Linux kernel and GPU firmware modification for the RTX 4090 (experimentally also for the RTX 4070 Ti SUPER and a few other Ada cards), Linux only. It touches the warranty and system stability on every boot, and it is built from source against DKMS and the kernel headers. It is not an install-and-forget tool for a regular user.
Open on GitHub▌ More finds