HP USB Disk Storage Format Tool: Simplifying USB Drive Management
Version: 2.2.3.1
License: free0
4.6 ( 8864 ratings )
OS: Windows
gpu-unlock-ada for RTX 4090: Linux Kernel Module for Tensor Cores
gpu-unlock-ada is an open-source Linux kernel module for GeForce RTX 4090 cards and, experimentally, the RTX 4070 Ti SUPER. It sets the streaming multiprocessor speed-select override before the stock NVIDIA driver starts, which raises the tensor-core instruction rate on consumer Ada GPUs. No patched NVIDIA driver and no UEFI application are needed.
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 release.
Key Facts
| Feature | Description |
|---|---|
| Purpose | Raises the tensor-core instruction rate of consumer Ada GPUs |
| Supported cards | GeForce RTX 4090; experimental support for the RTX 4070 Ti SUPER (device ID 10de:2705) |
| Platform | Linux only; kernel headers and build tools are required |
| Boot requirements | Cold boot; Secure Boot must be off |
| Firmware | The archive contains no NVIDIA firmware; it is generated locally from pinned NVIDIA inputs under NVIDIA's own license |
| Version | 1.1.1 (24.10.2026) |
| License | GPL-2.0 |
| Source | github.com/JustAResearcher/gpu-unlock-ada |
Reported Results
The release notes list the following instruction-throughput figures:
| Test | Before | After |
|---|---|---|
| RTX 4090, FP16 with FP32 accumulate | 165.6 TF/s | 331.6 TF/s |
| RTX 4090, BF16 with FP32 accumulate | — | 330.6 TF/s |
| RTX 4070 Ti SUPER, TF32 issue rate | 1.0× | 1.99× |
| RTX 4070 Ti SUPER, INT8, INT4, one FP8 GEMM | baseline | no improvement |
The authors state that these tests use register-resident loops and do not establish a speedup for real GEMM, inference or full models. One FP8 test reached 1.95× on a single card, but at lower core clocks than stock, and it covers only a narrow set of inputs. Independent cold-boot reproduction and numerical validation in applications were still pending when version 1.0.0 was published. Treat all numbers as author measurements, not Soft4Bro tests.
How to Install gpu-unlock-ada
- Download the source archive with the button above and check it against the SHA256SUMS file on the GitHub release page.
- Turn Secure Boot off and install the kernel headers and build tools.
- Follow the QUICKSTART in the repository to build the module and generate the firmware locally.
- Do a full cold boot, so the module can set the speed-select registers before the NVIDIA driver loads.
- Run a tensor benchmark to confirm the result and watch temperatures and power draw.
Before You Try It
| Advantages | Limitations |
|---|---|
| Doubles the measured FP16/BF16 tensor instruction rate on the RTX 4090 | Linux only, cold boot required, Secure Boot must be off |
| No BIOS flashing and no driver patching | Tested on a small number of cards |
| Open source code under GPL-2.0 | The RTX 4070 Ti SUPER support is experimental and limited to one device ID |
| Hash sums are published for the release files | Real application speedups are not proven; INT8 and INT4 did not improve |
Frequently Asked Questions
Does gpu-unlock-ada work on Windows?
No. It is a Linux kernel module.
Which cards does it support?
The main target is the GeForce RTX 4090. The RTX 4070 Ti SUPER is experimental. Other Ada cards are not officially supported.
Will it speed up LLM text generation?
Very little. Text generation depends mainly on memory bandwidth, not on tensor instruction rate. Training, fine-tuning and image generation have more to gain.
What about RTX 30-series cards?
Use gpu-unlock for RTX 30 Series.
Is it safe?
The tool changes GPU behavior at boot and is still new. Try it on a test machine, keep an eye on temperature and power, and use it at your own risk.
Conclusion
gpu-unlock-ada is an experimental Linux module for RTX 4090 owners who want more tensor throughput for training and image generation. Read the repository notes first, because the results so far come from a small number of tests.
