Independent guide · RTX 20 / 30 / 40
GPU Unlock
Your GeForce runs AI math at half speed. See if yours can run at full rate.
2×
FP16 / BF16 tensor rate
+25–55%
LLM prefill
0
driver patches or VBIOS flashes
01
Will it work on my GPU?
Pick your card, or paste the output of lspci -nn.
RTX 3090 — Yes. This card is on the stable list.
Measured BF16: 75 → 125 TFLOPS (+67%)
Install with: Zagitta/gpu-unlock
gpu_unlock.targets=all
A Windows UEFI port has reported success on this model.
Full details for RTX 3090 →02
Measured BF16 throughput
Dense matrix multiply, FP32 accumulation, stock power limits. Reported by the upstream project.
* RTX 4090 is the peak MMA rate; at stock power it gains much less in large GEMMs.
03
What gets faster
Faster
- LLM prefill (vLLM, PyTorch, FP16/BF16, AWQ/GPTQ)+25–55%
- Time to first token−25–30%
- LoRA and full fine-tuning in BF16+25–50%
- Image generation (SDXL, SD 1.5, DiT in ComfyUI)+30–55%
- Video generation (Wan 2.2, BF16)+15–20%
- FlashAttention / SDPA kernels+50–60%
Unchanged
- llama.cpp / Ollama with quantized GGUF0–15%
- Token generation (memory bound)0%
- INT8, FP8 and INT4 models0%
- Gaming and FP32 CUDA0%
04
How it works
Opt in at boot
You add one kernel option. Without it the module does nothing.
Unlock before the driver
A small module sets the tensor speed selectors, then resets the card.
Stock driver takes over
The unmodified NVIDIA driver loads as usual. A full power-off restores stock.
05
Pick the right project
Star counts on GitHub as of 2026-10-29.
Zagitta/gpu-unlock
The original module. Stable on RTX 30, experimental on RTX 20 and 40.
RTX 20 · 30 · 40 · Upstream
JustAResearcher/gpu-unlock-ada
RTX 4090 port with per-card opt-in flags and detailed evidence for other Ada cards.
RTX 4090 · RTX 40
impacifist/Windows-GPU-UNLOCK
UEFI port that unlocks before Windows boots. RTX 30, per-board profiles.
RTX 3060 – 3090 Ti
nlzy/turing-unlock
Turing adaptation, measured on the RTX 2080 Ti.
RTX 2080 Ti
physicsdolphin/nv-uefi-unlocker
UEFI unlocker in Rust for Windows, verified on RTX 20 and 40 cards.
RTX 20 · 40
phiraml/gpu-unlock-blackwell
Measures the cap on the RTX 5090 and maps out a port. No unlock yet.
RTX 5090
06
Stay safe
Build from the source repository. The upstream author has no website, coin or installer.
Be very careful with Windows installers that circulate outside GitHub.
Keep a stock boot entry, and cold-boot (power off) for the first try.
Check the kernel log for "gpu_unlock: unlocked" before you trust any benchmark.
07
FAQ
What is GPU Unlock?
gpu-unlock is an open-source Linux kernel module by Zagitta, published on October 18, 2026. GeForce cards run FP16, BF16 and TF32 tensor math with FP32 accumulation at half the rate of the professional cards built on the same silicon. The module removes that cap at boot, before the NVIDIA driver loads.
Does it damage my GPU?
The upstream author considers it very unlikely: normal throttling and power management stay in place. The card does more work, so it may run warmer and hit its power limit sooner. The Turing port warns that tensor cores draw a lot of power. Treat it as experimental.
Will it make games faster?
No. Games and FP32 CUDA do not use the capped path. Only FP16/BF16/TF32 tensor math with FP32 accumulation changes.
Will Ollama or llama.cpp get faster?
Barely. Quantized GGUF models gain 0–15%, mostly at long context, and token generation is limited by memory bandwidth. vLLM and PyTorch with FP16/BF16 weights see the big prefill gains.
Does it work on Windows?
Not from the upstream project, which is Linux only. Community UEFI ports run the unlock before Windows starts, but each one is tied to specific boards and cards. See the Windows page.
What about the RTX 50 series?
Not supported. A research project measured the same cap on an RTX 5090 (0.68× rate), but there is no public Blackwell payload and the upstream author calls a port unlikely.
Do I need a specific driver version?
No. The build downloads a pinned NVIDIA driver only to extract firmware. Your installed driver is not modified and nothing needs redoing after a driver update.
How do I undo it?
Remove gpu_unlock.targets= from the kernel command line, or delete it once in the bootloader edit screen. Removing the package removes the module. A full power-off returns the card to stock.
Is gpu-unlock.net the official site?
No. This is an independent guide. The upstream FAQ states the project has no official website. We only summarize public data and link to source code.
Ready to try it?
Find your card, then follow the install steps for your distribution.