GPU Unlock Spreads to RTX 40 Cards: RTX 4070 Ti SUPER Gets 2× Speed
The gpu-unlock kernel module that opened up restricted performance on RTX 30 cards is now expanding to RTX 40 series GPUs. The project has demonstrated an experimental unlock on the RTX 4070 Ti SUPER, with FP16 throughput increasing from approximately 95 TFLOPS to 190 TFLOPS, and BF16 showing nearly identical gains.
What happened
Developed by GitHub user Zagitta, gpu-unlock is an open-source Linux kernel module that enables higher Tensor Core instruction-issue rates on supported NVIDIA Ampere graphics cards. When it launched, it primarily targeted RTX 30 cards. Now, RTX 40-series unlocks are already emerging, with the Ada implementation currently targeting specific hardware, with the RTX 4070 Ti SUPER implementation still experimental.
An open-source Linux kernel module published on GitHub by developer Zagitta allows owners of older NVIDIA GeForce graphics cards to bypass software-level throughput restrictions on Tensor Core math, lifting half-precision compute performance by up to 80% without requiring a VBIOS flash or modified driver binaries. Rather than patching compiled graphics drivers after boot, the tool loads a custom kernel module via the initial RAM filesystem (initramfs) before the proprietary NVIDIA driver initializes. According to technical breakdowns, this early-boot module changes low-level hardware speed-selection registers. The driver binds normally afterward, never detecting a modified binary.
What people are building with it
On GitHub, the project is seeing active testing across card models and configurations. Compatibility reports show +83–84% GEMM improvements on an Inno3D RTX 4070 Ti, and +64–71% GEMM gains at 420W on RTX 3090 units. Creators running inference locally report substantial speedups in model serving. The unlock is most useful for dense tensor operations—the kind you're doing when running LLMs or image models on local hardware.
For a concrete example: One user confirmed that on a 3090 with a 360-watt power wall, throughput jumped from 75T to 135T when measured in Pytorch.
What's actually good
If you're running local models on a RTX 30 or (now) compatible RTX 40 card on Linux, this is a genuine free speed boost. The tool is easy to use and doesn't require any vBIOS flash or hardware modding. You're not voiding warranties or flashing firmware—you're loading a kernel module at boot.
Zagitta reports GEMM gains between +25% and +81% depending on card model and power configuration. The variance depends on which card you have, your power budget, and which operations you're running.
What isn't
The RTX 40 unlock is still experimental. Zagitta's original gpu-unlock remains primarily a Linux-based project, so Windows users aren't served. These types of "unlock" attempts are generally not supported by manufacturers and are risky. While performance gains are appealing, factors like stability, longevity, and security should not be overlooked. Before implementing such modifications in production environments or for critical workloads, extensive testing should be conducted, and potential risks carefully evaluated.
How to try it
Head to github.com/Zagitta/gpu-unlock. Install is a make install plus appending gpu_unlock.targets=all to kernel command line parameters. The GitHub repo has detailed docs for Debian, Fedora, Arch, and other distributions. Check the compatibility reports in the Issues section to see if your exact card model has been tested.
You'll need a system running Linux. The unlock doesn't affect your NVIDIA driver or require you to disable secure boot or patch binaries—it's remarkably clean as hardware hacks go.