Modders Unlocked Hidden AI Power on GeForce Cards. It Does Nothing for Emulation

Affiliate disclosure: This guide contains affiliate links. As an Amazon Associate, we earn from qualifying purchases at no extra cost to you. Learn more.

2026-10-26 · Explainer

Affiliate disclosure: This guide contains affiliate links. As an Amazon Associate, we earn from qualifying purchases at no extra cost to you. Learn more.

An open source project called GPU Unlock removes a software limit NVIDIA puts on GeForce cards, and roughly doubles how fast certain AI maths instructions can be issued. It went up on 18 October 2026 and collected more than 8,500 stars in a week.

Every headline says RTX 20, 30 and 40 series, and quotes an RTX 4090 going from 165 to 331 TFLOPS.

If you came here wondering whether this makes your emulators run better, the answer is no, and we will explain why in a moment. But two things in that paragraph above are wrong, and nobody is correcting them, so we may as well do both jobs at once.

The Short Version

It does not raise frame rates. Not in games, not in emulators. The project does not claim it does.
The real project covers the RTX 30 series only. Its own title says "Nvidia 3000's cards."
The 40 series claim rests on one machine. A side project written for one specific graphics card in one specific motherboard.
The viral TFLOPS figure comes from that side project, not from the main one, whose measured numbers are more modest.
What it removes is market segmentation, not a hidden turbo button.
The gains scale with available power. The weakest result came from power limited cards, which is the whole story for mobile hardware.
It helps emulation in exactly one indirect way, and that way was never short of performance.
No, it does nothing for the Switch 2, even though the Switch 2 uses the same GPU architecture.

Correction One: It Is Not Three Generations

The actual project targets four graphics chips: GA102, GA103, GA104 and GA106. Those are Ampere, and in consumer terms they are the RTX 30 series. The project's own one line description says it unlocks the instruction issue rate "on Nvidia 3000's cards to their professional equivalents."

The RTX 20 and RTX 40 claims come from separate repositories by other people, riding the same discovery:

RTX 40 series. A UEFI application written for, in its own description, one exact RTX 4090 model on one exact motherboard. That is a single machine proof of concept. It is not 40 series support.
RTX 20 series. A single compatibility report on one RTX 2070 Super, described by its own author as the first report for that chip.

So the honest version of the headline is "RTX 30 series, with early single machine reports on one card from each neighbouring generation." That is still interesting. It is not what anyone wrote.

Correction Two: The Big Number Is From the Side Project

The figure travelling fastest is an RTX 4090 going from 165.6 to 331.6 TFLOPS. That comes from the one machine 40 series effort.

The main project's own measured table is better documented and more modest. It is an RTX 3090 Ti at 450 watts, running a dense 8192 cubed matrix multiply, reported as five sample medians with FP32 accumulation.

Input	Before	After	Change
FP16	87.46 TFLOPS	118.39 TFLOPS	+35%
BF16	87.50 TFLOPS	157.41 TFLOPS	+80%
TF32	42.58 TFLOPS	82.61 TFLOPS	+94%

Worth noticing: the project says it doubles the rate at which these instructions can be issued, and only one of the three measured results comes close to doubling. Doubling the issue rate does not double the throughput, because something else becomes the limit. That gap between the theoretical change and the measured result is normal, and it is the part press coverage tends to drop.

Reported gains across the tested cards range from about +25% to about +84%, depending on the card and how it is configured.

What It Actually Removes

NVIDIA halves the rate at which GeForce cards can issue a particular family of instructions: FP16 and BF16 multiply accumulate where the result accumulates into FP32. The professional cards built on the same silicon do not have that halving.

That is product segmentation. It exists so that people doing serious compute work buy the expensive cards. The mod restores the rate the hardware is physically capable of.

Mechanically, it is a Linux kernel module that loads early in boot and writes some low level configuration values before NVIDIA's own driver takes control of the card. It is not a patched driver and not a modified firmware flash, which is why it survives driver updates. It also means:

It is opt in. The module does nothing unless you add a specific option to your kernel command line.
It is not permanent. Cutting power resets it. Suspending to RAM resets it too, and the module re-applies it on resume.
Hibernation is not covered, by the project's own admission.

We are explaining this rather than instructing it. This is a kernel module that resets your display card partway through boot, blanks the screen while it works, and whose documentation includes the line "power off before trying again" for when its firmware step fails to finish. If that reads as fine to you, you do not need a guide from us. If it does not, this is not for you.

There is also a licensing grey area the project notes itself: the payload it generates derives from an earlier reference, and NVIDIA's firmware "remains under its own terms." We will flag that and leave it there.

Why It Does Nothing for Games or Emulators

Three reasons, and they stack.

One: nothing in rendering uses that instruction path. Rasterising a frame is not matrix multiplication accumulating into FP32. The limit being lifted sits on a code path your emulator never executes.

Two: the project makes no gaming claim at all. Its documentation lists what gets faster: large matrix multiplies, AI model prefill, diffusion image generation. There is not a word about frame rates. When the people who built a mod do not claim a benefit, that is the strongest evidence available.

Three: even in AI work, it is selective. The project notes that the memory bandwidth bound half of running a language model sees no change, and neither does INT8 work. Outputs were bit identical before and after, which is the point, but it also tells you this is a narrow speedup rather than a general one.

If you want the upscaling that genuinely does raise frame rates in emulators, that is DLSS, FSR and XeSS, which is a different subject entirely. We cover it in is DLSS 5 worth it for emulation and FSR and XeSS on handhelds.

The Power Observation Is the Real Handheld Lesson

Here is the one finding in the project worth a handheld owner's attention, and it is a line most readers would skim past.

The project notes that the power limit matters a lot, and that the lowest gain it recorded, about +25%, came from cards capped at 225 watts. Cards with more power headroom gained far more.

The gain scales with the power you have spare. Which is precisely what mobile and handheld silicon does not have. A part running in a 15 to 30 watt envelope is already power limited at all times. Unlocking an instruction rate on a chip that cannot feed it more watts gets you much less than the headlines suggest.

Mobile Ampere does work, for what it is worth. The project's own table lists an RTX 3060 Laptop gaining 56 to 60%. A community report on a different 3060 laptop board claims around 70%. Some laptop boards only unlock when the driver is told to initialise the GPU earlier than usual, which appears to explain the scattered reports of no improvement at all.

And worth knowing: it does not even work on every Ampere card. One OEM RTX 3080 Ti variant completes the unlock successfully and its BF16 rate stays halved regardless, because that part is fused differently.

No, Not Even on the Switch 2

This is the question that occurred to us immediately, and it is a reasonable one, because the architectures match.

The Switch 2's T239 chip is Ampere, with some Ada Lovelace elements. It has 1,536 CUDA cores, 48 third generation tensor cores and 12 second generation ray tracing cores. Same generation as every card on that compatibility list.

The answer is still no, for concrete reasons:

It is not a discrete card sitting on a PCI bus, which is what the mod addresses.
There is no NVIDIA proprietary Linux driver loading after it, which is the window the mod operates in.
Most importantly, the limit being removed is a GeForce segmentation fuse. It exists to stop a gaming card competing with a professional card. A console chip has no professional equivalent to be protected from, so there is no equivalent limit to undo.

Nice hook, firm no. And while we are here: the Switch 2 remains the only NVIDIA powered gaming handheld worth mentioning, because no x86 handheld ships an NVIDIA GPU. Steam Deck, ROG Ally, Legion Go, the Odin line and the Retroid line are AMD, Qualcomm or MediaTek. There is no handheld in our catalog this mod could be pointed at.

Where It Touches Emulation At All

One place, and it is real, so it deserves a fair hearing.

HD texture packs for emulators like Dolphin and PCSX2 are made by running the game's original textures through AI upscaling models, usually from the ESRGAN family, through tools like Cupscale or chaiNNer. That work is FP16 convolution, which is exactly the path this mod speeds up.

So if you build texture packs, this could make your batch jobs finish faster.

Two caveats flatten it. First, that job was never short of performance. The existing guidance in that scene is that a GTX 1660 Super is already more than enough, and a 1660 Super has no tensor cores for this mod to unlock. Nobody was waiting on hardware. Second, the finished pack costs essentially nothing at runtime, because the work happened offline. Playing with an HD texture pack is just loading bigger images.

Net result: marginally faster to make a texture pack, zero difference to playing with one. Our full guide to what AI actually does for emulation, including the texture pack workflow, is AI and neural upscaling options that actually work.

The Bottom Line

This is a genuinely clever piece of work and a legitimately interesting story about how much performance sits behind a software flag on hardware people already own. If you run local AI models on an RTX 30 series card under Linux, go read about it properly.

For emulation and handheld gaming it changes nothing, and the three generation headline overstates what exists by two generations.

If you want more emulation performance, the levers that actually move are the ones we keep coming back to: a faster handheld, the right upscaler, and a sensible resolution.

Check Steam Deck OLED Price on Amazon

Frequently Asked Questions

Does the NVIDIA GPU unlock mod improve gaming performance?

No. It lifts a limit on a specific AI maths instruction path that rendering does not use. The project itself makes no gaming claim, and its documentation lists only AI workloads among the things that get faster.

Will the GPU unlock make my emulators run faster?

No. Emulators are limited by CPU work, graphics driver overhead and rasterisation, none of which touch the instruction path this mod unlocks. The one indirect benefit is faster AI upscaling when creating HD texture packs, and that task was never short of GPU power.

Which graphics cards does GPU Unlock actually support?

The RTX 30 series, specifically the GA102, GA103, GA104 and GA106 chips. Claims about the RTX 20 and 40 series come from separate side projects: one report on a single RTX 2070 Super, and a UEFI application written for one specific RTX 4090 in one specific motherboard.

Is the RTX 4090 really hitting 331 TFLOPS?

That figure comes from the single machine 40 series side project, not from the main one. The main project's own measured results on an RTX 3090 Ti are 118 TFLOPS FP16 and 157 TFLOPS BF16, up from about 87 TFLOPS for both.

Does the unlock work on laptop GPUs?

Yes, on Ampere laptop chips. The project's table lists an RTX 3060 Laptop gaining 56 to 60%, and a community report on another board claims around 70%. Some laptop boards need the driver told to initialise the GPU earlier, which appears to explain reports of no gain.

Can this be used on the Nintendo Switch 2?

No, despite the Switch 2 using the same Ampere architecture. It is not a discrete card on a PCI bus, there is no proprietary NVIDIA driver loading after it, and the limit being removed is a GeForce product segmentation fuse that has no equivalent on a console chip.

Does any gaming handheld have an NVIDIA GPU?

Only the Nintendo Switch and Switch 2. Every x86 handheld, including the Steam Deck, ROG Ally and Legion Go, uses AMD graphics, and Android handhelds use Qualcomm or MediaTek. There is no handheld this mod applies to.

Is the GPU unlock permanent?

No. It writes values that reset when the card loses power, so it is reapplied at every boot and again after waking from sleep. Hibernation is not covered. It is also opt in, and does nothing unless specifically enabled.

Is it safe to use?

It is a kernel module that resets your graphics card during boot and blanks the display while it works, and its own documentation tells you to power the machine off if its firmware step fails to complete. It also sits in a licensing grey area the project acknowledges. We are explaining what it does, not recommending it.

Written 26 October 2026. Project scope, chip list, measured figures, per card gain ranges, power observations and the licensing note come from the project's own documentation and its linked compatibility reports, retrieved that day. The RTX 20 and 40 series claims are attributed to separate repositories found the same day. Switch 2 chip specifications are as published by NVIDIA and Nintendo. We have not run this software and we publish no benchmark figures of our own.