gpu-unlock Explained: Up to 80% More FP16/BF16 Speed on RTX 3090/3080

gpu-unlock is a free, open-source Linux kernel module from developer Zagitta that removes Nvidia's artificial speed cap on FP16 and BF16 tensor math for GeForce RTX 30-series cards. It writes two hardware registers before the Nvidia driver loads, and on an RTX 3090 Ti that alone pushes BF16 throughput from about 87.5 TFLOPS to 157.4 TFLOPS, a jump of roughly 80%. The RTX 3090 gains about 65%, the RTX 3080 about 73%. It costs nothing, needs no VBIOS flash, and the catch is that none of it touches gaming frame rates.

If you've been running local LLMs, Stable Diffusion, or FLUX on a 3090 or 3080 and wondering why your card never felt as fast as its spec sheet promised, this is why. Nvidia has been quietly running half-rate FP16/BF16 math on consumer Ampere cards since launch, reserving the full rate for the A100 and RTX A6000. gpu-unlock just flips that switch back on.

What is gpu-unlock?

gpu-unlock is a small kernel module, built and maintained by a developer going by Zagitta on GitHub, that targets a specific limitation on Nvidia's consumer Ampere GPUs: FP16 and BF16 tensor instructions with FP32 accumulation run at half the instruction-issue rate compared to Nvidia's professional cards built on the same silicon. That's the exact math most local AI training, fine-tuning, and unquantized inference actually uses.

The project isn't a driver patch, a VBIOS mod, or a jailbreak. It's a module that runs once at boot, flips a couple of registers, and gets out of the way before the stock Nvidia driver even loads. Reboot without it and the card goes right back to factory behavior. The repo passed 5,000 GitHub stars within days of going public in October 2026, and it's since spawned experimental ports for RTX 20 and RTX 40 cards too, though RTX 30 is where the numbers are most consistent.

How much faster is it, really?

The gains depend entirely on which number you're looking at: raw tensor math, or a real application running on top of it. Here's what the project's own benchmarks show for dense BF16 matrix multiplication:

TF32 math gains even more on some cards, reportedly jumping close to 94% on the 3090 Ti. But raw TFLOPS isn't what you run day to day, so people started pointing real workloads at it. One tester's FLUX image generation dropped from 32 seconds to 21 seconds per image. ComfyUI runs reported going from about 72 seconds to 45 seconds. vLLM prompt prefill, the step where the model reads your input before it starts generating, came in anywhere from 25% to 55% faster depending on sequence length and batch size.

Token generation is the odd one out. Several testers reported "barely changed" decode speed, with one post putting the gain at only around 16% during actual LLM inference. That's not a bug in the mod, it's just physics: decode is memory-bandwidth bound, not compute bound, so doubling your tensor math doesn't do much when the bottleneck is how fast data moves off VRAM.

How does gpu-unlock actually work?

The technique is a lot less dramatic than "doubling your GPU," which is honestly refreshing in a space full of overclaiming. Here's the short version:

1. A kernel module loads during early boot, via initramfs, before the standard Nvidia driver binds to the card.
2. It writes to two hardware registers Nvidia calls Speed Selectors, SS0 and SS1, that gate how fast the tensor cores issue FP16/BF16/TF32 instructions.
3. It performs a PCIe Function Level Reset on the GPU.
4. It steps aside and lets the completely unmodified, stock Nvidia driver load normally.

No VBIOS flash. No patched driver binary. No UEFI app required. That's the main reason it's spread so fast: the risk profile is a lot lower than the shunt-mod or BIOS-flash unlocks GPU miners used to pull on these same cards years ago. Reboot, and you're back to stock. The output math is bit-identical too, it's purely an issue-rate change, not a precision shortcut.

Which GPUs does it support?

As of the version referenced here, support breaks down roughly like this:

• Stable, well-tested: RTX 30-series (Ampere), including the 3050, 3060, 3060 Ti, 3070, 3070 Ti, 3080, 3080 Ti, 3090, 3090 Ti, and most laptop variants. The underlying dies are GA102, GA103, GA104, and GA106.
• Experimental: RTX 20-series (Turing) and RTX 40-series (Ada Lovelace), both 60 and up through the top-end cards.
• Not supported: RTX 50-series (Blackwell). Nvidia changed enough on that architecture that the current approach doesn't carry over.
• Platform: Linux only, built on top of the mainline Nvidia driver. A separate community Windows port exists but it's rougher and less documented. If you're on Windows for your AI box, you're waiting on that project to catch up, not this one.

How do you install gpu-unlock on Linux?

If you're comfortable editing your bootloader config, this is a short list:

1. Clone the repo and build it: git clone https://github.com/Zagitta/gpu-unlock && cd gpu-unlock && make install
2. Add gpu_unlock.targets=all to your kernel command line. On GRUB-based distros (Debian, Ubuntu, Arch), edit GRUB_CMDLINE_LINUX_DEFAULT in /etc/default/grub, then run sudo update-grub (or grub-mkconfig -o /boot/grub/grub.cfg on Arch). On Fedora or RHEL, use sudo grubby --update-kernel=ALL --args="gpu_unlock.targets=all". On systemd-boot, append it to the options line in your boot entry.
3. Reboot.
4. Check that it took effect, either with dmesg output from the module or by running a quick BF16 matmul benchmark in PyTorch before and after.

That's genuinely it. No firmware tools, no soldering, no shunt resistors.

What won't gpu-unlock speed up?

This is the part a lot of coverage glossed over, so it's worth being blunt: this is a tensor-core unlock, not a GPU overclock.

• Gaming frame rates. No independently verified benchmark has shown an FPS gain from this. DLSS uses tensor cores too, but nobody's published a solid before/after on actual games yet. Don't install this expecting more frames in your favorite title.
• Token generation / decode speed in LLMs. As noted above, decode is bandwidth-bound. You'll see it in prefill and fine-tuning, not in how fast tokens stream out during chat.
• Anything that isn't FP16/BF16/TF32 tensor math. INT8 inference, CUDA core workloads, video encode/decode, none of that touches the registers this mod changes.
• VRAM capacity or bandwidth. The mod doesn't add memory or bandwidth. If your workload was VRAM-limited before, it still is.

Is it safe to run on your GPU?

For a mod that changes hardware registers, the risk profile is genuinely mild, but "mild" isn't "zero." A few specifics worth knowing before you flip it on:

• It's reversible by reboot. There's no persistent firmware flash, so you can't brick the card this way the way old VBIOS mods could.
• Thermal and power limits stay exactly where Nvidia set them, the mod only removes the artificial instruction-rate cap, not any safety throttling.
• Laptops are the weak spot. If the dedicated GPU powers down during idle or sleep, the unlock silently resets and you're back to stock until next boot.
• Suspend/resume gets slightly slower, by about a second, and hibernation isn't supported yet.
• It's still labeled experimental by its own author for anything outside stable Ampere. Test on a secondary machine before you put it on a daily-driver workstation, and don't assume it's identical across driver versions.
• The license is GPL-2.0 for the module itself; the generated firmware traces back to Nvidia's own CMP reference code and inherits Nvidia's terms there. This sits in enough of a legal gray zone that I wouldn't put it anywhere near a production server without checking your support contract first.

Should you actually install this?

If you're running a 3090 or 3080 for local LLM fine-tuning, FLUX, or Stable Diffusion on Linux and you're comfortable editing a GRUB config, I'd say yes, try it. The downside risk is low, it's free, and a 65-80% jump in the exact math local AI tooling leans on is not a rounding error. I'd stop short of calling it a free upgrade for everyone, though. If most of your workload is chatbot-style token generation, you're not going to feel much of this, and if you're gaming on the same box, manage your expectations: this isn't an FPS unlock no matter how many headlines frame it that way.

The bigger story here isn't really one GPU mod. It's that Nvidia has been segmenting consumer and professional cards with a software switch rather than a hardware difference, and now that switch is public. Expect more scrutiny of similar caps on future cards, and don't be surprised if Nvidia closes this specific door in a firmware update somewhere down the line.

If you're picking tools for a local AI setup more broadly, our roundup of the best AI tools in 2026 covers the software side of that stack. And if you're using the extra throughput to fine-tune reasoning models locally, Sebastian Raschka's approach in building reasoning models from scratch is a good next read. For image generation specifically, the diffusion gains people are reporting pair well with what we covered in Mercury 3's diffusion LLM approach.

Watch and read more