GPU Unlock: A Free Speed Boost for RTX 30-Series Cards Running AI Image and Video Models, and the Test to Run Before You Trust It
A free Linux add-on lets RTX 3060 to 3090 Ti cards do the maths most image and video models use at full speed. One owner measured close to 60 percent faster renders with identical pixels. Here is who should try it, and how to check it on your own work.
The RTX 3090 in a lot of home studios has spent its whole life doing one kind of maths at half the speed its chip can manage. That was a product decision, not a physical limit. NVIDIA sells cards built on the same family of chips to professionals, and the cheaper gaming versions run certain number formats at a reduced rate.
Those number formats are the ones most modern image and video models use. On October 18 a developer called Zagitta published a free project, GPU Unlock, that sets those cards back to the professional rate when Linux starts. It has collected about 8,000 GitHub stars in eight days, which is a lot of people with old graphics cards suddenly paying attention.
The headline figure is from the project's own benchmark on an RTX 3090 Ti: one type of maths goes from 87.5 to 157.4 trillion operations per second. That number means very little to someone who makes posters or short films. The number that matters is wall-clock time on a real render, and that one is more interesting, and more uneven, than the headline.
What changed, in plain terms
Most current image and video models store their numbers in a format called BF16, a 16-bit way of writing decimals that keeps file sizes manageable. Your card does millions of sums in that format for every frame. On a GeForce 30-series card those sums have been rationed.
GPU Unlock is a kernel module, which means a small piece of system software that loads while Linux boots. It runs before NVIDIA's normal driver takes over the card, sets the card's speed setting, then resets the card and hands it to the ordinary driver. Because it never touches the driver itself, a driver update does not undo it.
The project lists user reports for the RTX 3090 Ti, 3090, both 3080 models, the 3070 and the 3060 12 GB. There are no reports yet for one chip variant, and one OEM 3080 Ti completed the process but stayed at half speed. If your card is not on that list, nobody has shown it works.
The real-world numbers, and the twist inside them
ComfyUI Wiki collected one 3090 Ti owner's before-and-after timings in ComfyUI, using the median time for each model. This is the most useful evidence so far, and also a single person on a single card, so read it as a strong hint rather than a promise.
In the full-size files, the gains were large:
- MOSS-Image, the free poster model: 58.6 percent faster
- MiniMax H3 video with sound: 60.2 percent faster
- Qwen-Image 2.1 Turbo: 57.2 percent faster
- Krea 2 Turbo: 46.3 percent faster
- LTX-2.5's distilled (speed-trimmed) version: 16.8 percent faster
The same owner matched 150 still images from before and after, and the decoded pixels were identical. The card got faster without changing your picture. For anyone delivering client work, that is the property that makes this worth considering at all.
Now the twist. Many people with 24 GB cards run compressed 8-bit versions of these models, files that store each number in half the space so the model fits in memory and runs faster. Those barely moved. The 8-bit MOSS-Image file gained 0.3 percent. The 8-bit Qwen-Image file gained 4.6 percent. One 8-bit H3 file did gain 31.2 percent, so there is no clean rule, but the pattern leans one way.
That matters for a reason the launch chatter skipped. For the past year, the compressed file was the speed choice. You accepted a small quality hit to wait less. If the full-size file just got 58 percent faster on your card and the compressed one did not, the gap between them shrinks, and the full-size file may now be the better trade if it fits in your memory. The test below is how you find out for your own setup, instead of guessing.
Who should try this, and who should not
Try it if all of these are true: you already run Linux, you own one of the listed 30-series cards, and you render with full-size model files often enough that a faster render changes your day. A motion designer running MiniMax H3 shots overnight, or an illustrator iterating on dozens of MOSS-Image drafts, gets real time back.
Skip it if you are on Windows or a Mac. The project is Linux only. A separate community port for Windows exists, but ComfyUI Wiki did not test it and neither has this article. Skip it too if you mostly run compressed 8-bit files; your gain may round to nothing. And skip it if the machine is your only work computer on a deadline week. This software touches your graphics card at a low level, and ComfyUI Wiki's own advice is to "treat it as experimental."
Do not buy a used 3090 because of this. The speed-up does not add memory, and memory still decides which models you can run at all.
Put this into practice: the same-seed test
The whole point of a test is to measure your models, on your card, with your settings. It takes about an hour, most of it waiting.
1. Record a baseline before you install anything. In ComfyUI, pick the workflow you use most, with a full-size model file. Fix the seed (the number that decides the random starting point), keep the prompt and settings identical, and render five images or five short clips. Write down each time and take the middle one. Save the outputs in a folder called "before."
2. Add a second baseline with a compressed file, if you have one. Same prompt, same seed, five renders. This is what lets you see whether the twist above applies to you.
3. Install the module. The README gives one command per Linux family: make install, or make install-deb, make install-rpm or make install-arch. Then add gpu_unlock.targets=all to your kernel command line (the README shows how for GRUB, Fedora's grubby and systemd-boot). The important step people skip: power the machine fully off, then start it. A warm restart is not the same.
4. Confirm it took. Run journalctl -k -b | grep gpu_unlock and look for the module's success lines. If your screen goes blank for a moment during boot, that is expected; the card resets before the normal driver takes over.
5. Rerun both tests exactly. Same workflow, same seed, five renders each. Note the middle times. Save the outputs in "after."
6. Check the pixels. Open a before and after image as two layers in Photoshop, Affinity or GIMP and set the top layer to Difference. A pure black result means identical pixels. For video, do the same with one frame from the middle of a clip.
7. Decide. If the full-size file is now meaningfully faster and fits in memory, consider making it your default. If nothing moved, remove the setting and move on.
To undo it, delete gpu_unlock.targets= from the kernel command line, or remove the package, which also removes the module. For a one-time test boot you can delete the setting from the bootloader's edit screen and leave the system untouched.
The honest limitations
It is one person's test so far. The ComfyUI timings come from a single 3090 Ti owner. Power limits change everything; the README's lowest reported gain (25 percent) came from cards capped at 225 watts, and it says the power limit "matters a lot." Your card, cooling and power setting will give your own number.
Some files will not get faster. The gain lands mostly in full-size files. If your workflow is built on compressed 8-bit files, the benefit may be close to zero.
It is low-level software from a new project. ComfyUI Wiki describes it as an unsigned kernel module. The README notes that if a card's firmware does not respond, you need to power off before trying again. Hibernation is not covered, and sleep needs a short re-run on wake, which the module handles by default. Back up and give yourself a quiet afternoon.
It does not fix memory. A model that did not fit on your card yesterday still does not fit today. Faster is not bigger.
The 4090 version is a different story. A fork for the RTX 4090, gpu-unlock-ada, appeared on October 23. Its README reports roughly double the speed in synthetic instruction tests, but only about 12 to 34 percent in matrix-maths tests in one owner's run, with one setting unchanged at one size. It says these are not whole-app speedups, and an unlocked 4090 peaked at 346 watts against 313 stock. Treat it as earlier and less proven than the 30-series module.
The bigger point
Creative AI coverage tends to measure progress in new models. This week the more useful upgrade for a lot of home studios was older hardware doing work it was always capable of. That is worth noticing, because it rewards people who keep their tools local: no subscription went up, no model was retired, and the pictures came out the same.
It also tells you something about choosing files. "Use the compressed version for speed" was a rule of thumb built on a limit that, for some of you, no longer exists. Rules of thumb in this field have a short shelf life.
So run the test. If your card is on the list and your workflow uses full-size files, you might get the best kind of upgrade: the same picture, sooner, for nothing. If the numbers do not move, you lost an hour and learned exactly what your machine is good at. Either way, you will be deciding from your own renders instead of someone else's benchmark.