PhoenixMiner: The Ultimate ETH Mining Software for Maximum Profitability
Version: v6.2c
License: free0
4.6 ( 8864 ratings )
OS: WindowsLinux
gpu-unlock for RTX 30 Series: Linux Kernel Module for Tensor Math
gpu-unlock is an open-source Linux kernel module that raises the FP16 and BF16 instruction issue rate of GeForce RTX 30-series cards (GA102, GA103, GA104 and GA106) to the level of their professional equivalents. It loads early in the boot process, before the NVIDIA driver takes control of the card, and changes the card's speed selectors. It is neither a patched NVIDIA driver nor a UEFI application.
The project is made by the community. It is unofficial and is not affiliated with NVIDIA. Soft4Bro lists it for information and links to the original GitHub repository.
Key Facts
| Feature | Description |
|---|---|
| Purpose | Unlocks the FP16/BF16 FMA instruction rate with FP32 accumulation on GeForce 30-series GPUs |
| Supported chips | GA102, GA103, GA104, GA106 (RTX 30 series) |
| Platform | Linux only; kernel headers and build tools are required |
| How it works | Runs before the NVIDIA driver binds the card, writes the speed selectors and resets the GPU |
| Installation | Build from source with make install (targets for Arch, Debian/Ubuntu and RPM-based systems are included) |
| Activation | Opt-in through the kernel parameter gpu_unlock.targets |
| License | The repository does not specify a license |
| Source | github.com/Zagitta/gpu-unlock |
Reported Results
The project README gives the following figures for an RTX 3090 Ti at 450 W (8192 × 8192 × 8192 dense GEMM, FP32 accumulation):
| Data type | Before, TFLOPS | After, TFLOPS |
|---|---|---|
| FP16 | 87.5 | 118.4 |
| BF16 | 87.5 | 157.4 |
| TF32 | 42.6 | 82.6 |
Users report BF16 matrix multiplication gains from about 25% to 81% depending on the card and its power limit. LLM prompt processing (prefill) improves by roughly 14–36%. LLM text generation (decode) and INT8 workloads stay about the same, because they depend mostly on memory bandwidth. These numbers come from the project authors and community reports, not from Soft4Bro tests.
Cards Reported to Work
- RTX 3090 Ti and RTX 3090
- RTX 3080 10 GB and RTX 3080 20 GB
- RTX 3070
- RTX 3060 12 GB
On the RTX 3080 Ti 20 GB OEM card the unlock completes, but BF16 stays at half rate. The module has not been tested on the RTX 3070 Ti. There are no reports for GA103 cards yet.
How to Install gpu-unlock
- Download the source archive with the button above or clone the GitHub repository.
- Install the kernel headers and build tools for your distribution.
- Run
make install, or usemake install-arch,make install-debormake install-rpm. - Add
gpu_unlock.targets=all(or specific PCI addresses) to the kernel command line. - Update the initramfs with the tool your distribution uses, so the module loads at boot.
- Reboot and run a matrix multiplication benchmark to confirm the new speed.
The build downloads pinned inputs from NVIDIA and CachyOS and checks them with SHA-256, so network access is needed during installation.
Before You Try It
| Advantages | Limitations |
|---|---|
| Higher FP16/BF16 matrix throughput on supported cards | Linux only |
| No driver patching and no BIOS flashing | The display card shows a blank screen from the reset until nvidia-drm takes over |
| Opt-in: the module does nothing without the kernel parameter | Cards already bound to a driver are skipped |
| Re-unlocks the card after suspend to RAM | If the card firmware does not stop, a full power-off is needed before retrying |
| Open source code | Unofficial; no warranty and no independent long-term stability data |
Frequently Asked Questions
Does gpu-unlock work on Windows?
No. This module is written for Linux. A Windows route is not part of this download.
Will it make local LLM chat faster?
Only a little. Prompt processing can speed up, but text generation depends on memory bandwidth, so the gain is small.
Is it good for training and image generation?
Matrix-heavy work such as neural network training, LoRA fine-tuning and image or video generation can benefit. The real gain depends on the model, the software and the card's power limit.
Does it support RTX 40-series cards?
This project covers the 30 series. For the RTX 4090 see gpu-unlock-ada.
Is it safe?
The tool changes how the GPU behaves at boot. It is experimental community software, so try it on a test machine first and use it at your own risk.
Conclusion
gpu-unlock is a small Linux module for owners of RTX 30-series cards who run training or image generation and want more FP16/BF16 throughput. Read the README, check that your card is on the supported list and keep a way to boot without the module.
