RTX 30 GPU Unlock Mod: Up to 80% Faster AI Maths on Older Cards, and Is It Safe?

An open source project called GPU Unlock is trending this week because it makes older NVIDIA GeForce RTX 30 cards much faster at certain AI tasks. Early reports show gains of up to 80% in one benchmark on an RTX 3090 Ti, with no new hardware. However, the headlines skip the fine print. This RTX 30 GPU unlock does not boost gaming, works mainly on Linux and is still experimental.

Here is what it does, the real numbers, and whether you should try it.

What the RTX 30 GPU unlock mod actually does

GeForce RTX 30 cards use NVIDIA’s Ampere chips. On these gaming cards, NVIDIA runs one kind of AI maths at half speed: FP16 and BF16 calculations that store results in FP32. Professional Ampere cards with the same chips run that maths at full speed.

GPU Unlock, released by GitHub user Zagitta, removes that limit. According to its README, it “doubles the instruction issuing speed” of these operations, so a GeForce card behaves like its professional sibling for this one workload.

This type of maths powers a lot of modern AI. For example, image models such as Stable Diffusion style tools and Qwen Image run in BF16. As a result, they gain the most.

How the RTX 30 GPU unlock works

The method is clever and fairly clean. GPU Unlock is a small Linux kernel module that loads before the NVIDIA driver. It then:

1. Writes new values to the card’s “speed selector” registers.
2. Resets the GPU so the new setting takes effect.
3. Hands over to the normal, unmodified NVIDIA driver.

In other words, it does not flash the card’s BIOS, patch the driver or fake the card’s identity. The project says it does not need changes for each new driver release. In addition, it is opt in, so nothing happens until you add a setting to your boot options.

RTX 30 GPU unlock benchmark gains

Every figure so far comes from the developer or users, not from independent lab tests. Still, the pattern is consistent.

The best known test comes from the README. On an RTX 3090 Ti at 450 W, BF16 throughput rose from 87.5 to 157.4 TFLOPS. FP16 rose from 87.5 to 118.4 TFLOPS.

Real apps gain less than raw maths tests. A Reddit user running ComfyUI on an RTX 3090 Ti reported image models about 56 to 60% faster. Meanwhile, INT8 models barely changed.

What about LLMs and gaming?

For local chatbots, the picture is mixed. Prompt processing got roughly 14 to 36% faster in reported tests. However, the speed of generating words barely moved, because memory bandwidth, not maths, limits that step. So if you run large language models at home, do not expect a big jump. Our guide to running AI locally explains why memory matters more there.

For games, nobody has shown higher frame rates. DLSS uses the same Tensor Cores, so a small gain is possible in theory. Still, there is no proof yet. Note that separate “DLSS frame generation unlocker” mods for RTX 20 and 30 cards are different projects, and they are not part of GPU Unlock.

Which cards the RTX 30 GPU unlock supports

Reports cover cards with the GA102, GA104 and GA106 chips. That includes most RTX 3060 12GB, 3070, 3080, 3090 and 3090 Ti models. There are no reports yet for GA103 chips, and an OEM RTX 3080 Ti 20GB stayed at half rate in one test.

The main project does not support RTX 40 cards. A separate experimental fork for the RTX 4090 exists, but its own author says the numbers are raw issue rates, not real app speedups.

Does it work on Windows?

Not officially. The main project is Linux only. However, a separate project called rtx30 unlock applies the unlock with an EFI bootloader before Windows starts. Its first release arrived on 21 October 2026, and the author tested it on one RTX 3090 Ti. The author warns that it may only work on that specific setup, so treat it as very early.

Is the RTX 30 GPU unlock safe?

It is experimental, so weigh these risks first:

• Stability: after the reset, the screen stays blank until the driver loads. If something goes wrong, you may need a hard power off.
• Unsigned boot code: the Windows route runs code before your operating system loads. Only use code you have reviewed or built yourself.
• Warranty and terms: no source has confirmed whether it affects your warranty or NVIDIA’s terms. NVIDIA has not commented so far.
• Power and heat: faster maths can mean more power draw. Keep an eye on temperatures and power limits.

Because the mod flashes nothing, removing the boot setting returns the card to normal. Even so, keep a standard boot entry ready, as the RTX 4090 fork advises.

Should you try it?

If you run Stable Diffusion style image models, LoRA training or other BF16 work on Linux with an RTX 3080, 3090 or 3060 12GB, this mod could save real time. Test it on a spare boot entry, compare speeds, and check stability under long loads.

On the other hand, gamers and Windows users should wait. The Windows route is too new, and there is no gaming benefit. If you plan a new machine instead, our guides to the best PC for AI and NPU vs GPU can help.

One more effect is worth noting. Used RTX 3090 prices have already risen sharply this year due to AI demand. News like this may push them higher, so check prices before you buy a used card for this mod.

FAQ

Will the RTX 30 GPU unlock increase gaming FPS?

No gains have been shown. The mod speeds up FP16 and BF16 Tensor Core maths used in AI work, not normal game rendering.

Does it work on Windows?

The main project is Linux only. An experimental EFI loader called rtx30 unlock brings it to Windows, but its author has only tested it on one RTX 3090 Ti.

Do I need to flash my BIOS or patch the driver?

No. It is a kernel module that changes a register setting before the standard NVIDIA driver loads.

Which RTX 30 cards work?

Most cards with GA102, GA104 and GA106 chips, from the RTX 3060 12GB up to the RTX 3090 Ti. GA103 cards have no reports yet.

Is the mod permanent?

No. The mod changes a setting that resets when the card powers off, and nothing is flashed to the card. If you remove the boot setting, the card runs at its normal speed again.