GPU Unlock: full-rate FP16 and BF16 on RTX 3000 and 4090
GPU Unlock is a free, open-source Linux kernel module that doubles the FP16/BF16 tensor issue rate on NVIDIA RTX 3000 cards such as the RTX 3090 and 3080. A fork extends it to the RTX 4090. This unofficial guide explains what it does, which GPUs work, how to install it and what the risks are.
GPU Unlock 1.0.0 (RTX 3000 series, by Zagitta) and GPU Unlock Ada 1.1.1 (RTX 4090 and experimental RTX 40 cards, by JustAResearcher). Linux only.
GPU Unlock on GitHubGPU Unlock Ada (RTX 4090)How to install
Both buttons open the official GitHub repositories. This site does not host any files.
What GPU Unlock does
GeForce RTX 3000 cards run some tensor instructions at half the rate that NVIDIA’s professional cards with the same chips use. Two hidden registers, the SM speed selectors, set that rate. They cannot normally be written from the driver. GPU Unlock uses a known flaw in the card’s boot firmware, found in earlier work on NVIDIA’s CMP mining cards, to make those registers writable. It then sets every selector to full rate and resets the card before the normal NVIDIA driver starts.
The result, according to the project, is that FP16 and BF16 fused multiply-adds into FP32 issue twice as fast on GA102, GA103, GA104 and GA106 chips. That is the whole RTX 3060 to 3090 Ti range.
How much faster is it?
| Input (FP32 accumulate) | Before | After |
|---|---|---|
| FP16 | 87.46 TFLOPS | 118.39 TFLOPS |
| BF16 | 87.50 TFLOPS | 157.41 TFLOPS |
| TF32 | 42.58 TFLOPS | 82.61 TFLOPS |
RTX 3090 Ti, dense 8192³ GEMM at 450 W, median of five runs, from the first report on the project's issue tracker.
Other owners report BF16 GEMM gains from about +25% (cards capped at 225 W) to about +80%. In real workloads, LLM prefill got 14–36% faster in four reports, and diffusion and FlashAttention also speed up. LLM decode, which is limited by memory bandwidth, and INT8 work stay about the same. Reports that compared outputs found them bit-identical before and after.
GPU Unlock at a glance
Which GPUs work?
Owners have reported success on the RTX 3090 Ti, 3090, 3080 10 GB, 3080 20 GB, 3070 and 3060 12 GB. The OEM RTX 3080 Ti 20 GB completes the unlock but BF16 stays at half rate. The RTX 4090 works with the Ada fork. The full list with PCI IDs is on the supported GPUs page.
How to install it, in short
git clone https://github.com/Zagitta/gpu-unlock
cd gpu-unlock
make installThen add gpu_unlock.targets=all to your kernel command line and reboot. The module does nothing until that option is set. The install guide covers GRUB, Fedora, systemd-boot and how to check the result. For an RTX 4090 use the Ada fork instead; see GPU Unlock for RTX 4090.
Things to know first
- It is unofficial. NVIDIA does not support it, and it relies on a firmware flaw NVIDIA could close in future cards or firmware.
- Linux only. It needs kernel headers, DKMS and an initramfs generator. See GPU Unlock on Windows.
- It runs at every boot. The setting is lost when the card powers off, so the module runs again on each boot and after suspend to RAM.
- Read the risks. It runs firmware code on your GPU before the driver. See is GPU Unlock safe.
GPU Unlock FAQ
What is GPU Unlock?
GPU Unlock is an open-source Linux kernel module that unlocks the full FP16, BF16 and TF32 tensor issue rate on NVIDIA GeForce RTX 3000 cards, the rate their professional counterparts run at. A fork called GPU Unlock Ada adds the RTX 4090 and, experimentally, the RTX 4070 Ti SUPER and RTX 4060 Ti 16GB.
Is GPU Unlock made by NVIDIA?
No. It is an independent project on GitHub by Zagitta, built on earlier community research into the CMP mining cards. NVIDIA does not make, support or endorse it.
Does GPU Unlock make games faster?
Not in any way the projects measure. The gain is in FP16, BF16 and TF32 matrix math with FP32 accumulation, the kind used for AI training, LLM prompt processing, diffusion models and FlashAttention. Ordinary FP32, INT8 and LLM token generation stay about the same.
Does GPU Unlock work on Windows?
No. Both projects are Linux kernel modules. There are separate experimental Windows UEFI experiments, but no Windows installer. See the Windows page.
Do I need a patched NVIDIA driver?
No. The module runs before the stock NVIDIA driver loads, sets the speed selectors and resets the card, then the unmodified driver binds as usual. You do not need to re-patch anything when NVIDIA releases a new driver.
How do I turn GPU Unlock off?
Remove gpu_unlock.targets= from the kernel command line. For a single boot you can delete it in the bootloader editor. Removing the package also removes the module.
Is GPU Unlock free?
Yes. The kernel module and packaging are released under GPL-2.0-only. NVIDIA firmware stays under NVIDIA’s own terms and is generated on your machine during the build.