A new open-source Linux kernel module called gpu-unlock says it can double one of the most important AI math speeds on Nvidia’s RTX 3000 series cards. It lifts the FP16 and BF16 instruction issue rate on GeForce Ampere GPUs to match their professional cousins. According to Hacker News, where the project picked up 161 points, the tool runs before Nvidia’s own driver loads. You keep the stock driver and don’t patch anything.
Why does that matter? Nvidia has long run consumer GeForce cards at half speed for tensor operations that use FP32 accumulation, while its pricier workstation cards with the same silicon run at full speed. gpu-unlock removes that gap on GA102, GA103, GA104 and GA106 chips. That covers everything from the RTX 3060 to the 3090 Ti.
The numbers are big
The project’s own test used an RTX 3090 Ti at 450 W, running a dense 8192³ matrix multiply with FP32 accumulation. Each figure is the median of five runs:
- FP16: 87.46 to 118.39 TFLOPS
- BF16: 87.50 to 157.41 TFLOPS (about 80% faster)
- TF32: 42.58 to 82.61 TFLOPS (nearly double)
TFLOPS means trillions of floating-point operations per second. For dense matrix math, which is the core of neural network work, you’re getting close to a free generational jump.
How it works
The module takes the V67/PLM technique from the earlier CMP unlocker and pairs it with the early timing of a UEFI-based approach. It runs the unlock, writes the speed selector registers, and resets the card (a function-level reset, or FLR) before the unmodified Nvidia driver grabs it.
What stands out here is the maintenance story. Older methods meant re-patching every time Nvidia shipped a new driver. This one sits in the initramfs and loads before the driver, so driver updates shouldn’t break it.
Community reports are mostly positive
Users have posted compatibility reports for a range of cards, with BF16 GEMM gains of:
- RTX 3090 Ti: +68 to 80%
- RTX 3090: +25 to 67% (seven reports)
- RTX 3080 10 GB: +70 to 73%
- RTX 3080 20 GB (memory-upgraded): +56 to 59%
- RTX 3070: +73 to 81%
- RTX 3060 12 GB: about +80%
Reports also cover rigs with up to four cards, the card driving the display, and GPUs passed through to a virtual machine with vfio-pci.
Real AI workloads speed up, but not all of them
Matrix-heavy jobs benefit most. Users saw LLM prefill (the prompt-processing stage) get 14 to 36% faster, and diffusion models and FlashAttention improved too.
LLM decode, the token-by-token generation stage, stays about the same. That stage is limited by memory bandwidth, not compute. INT8 workloads don’t change either. Reports that compared outputs found them bit-identical before and after, so accuracy isn’t affected.
The caveats are real
This isn’t a one-click install for everyone. Keep these in mind:
- Linux only. It’s a kernel module that you opt into through the kernel command line (
gpu_unlock.targets=allor specific PCI addresses). - Not every card works. On an OEM RTX 3080 Ti 20 GB, the unlock finishes but BF16 stays at half rate.
- Gaps in testing. No one has reported on GA103 cards yet, and the 3070 Ti has only been tested through a UEFI port on Windows.
- Power limits matter. The weakest result (+25%) came from cards capped at 225 W.
- Benchmarks vary. Each user ran their own test, so you can’t compare the gains directly.
On the plus side, installation is clean for an unofficial tool. make install detects Arch, Debian/Ubuntu or Fedora/RHEL and builds a native package. Builds pull pinned Nvidia inputs over HTTPS and check their SHA-256 hashes.
Why this matters
The used RTX 3090, with 24 GB of VRAM at a fraction of the price of new hardware, is already a favorite for local AI. An unlock that pushes its BF16 throughput toward workstation levels makes it even better value for fine-tuning, image generation and long-context prompt processing.
It’s also a reminder of how much of GPU product segmentation comes down to firmware and driver choices, not silicon. Nvidia hasn’t publicly responded, and it’s fair to wonder whether future firmware will try to close this path. For now, anyone running Ampere cards on Linux has a reason to try it. You can find full install steps and the compatibility tracker on the original Hacker News thread.