Is Nvidia secretly throttling your GPU’s AI power? A free Linux tool claims to unlock it

Two open-source Linux kernel modules are claiming to strip a speed limit out of Nvidia’s consumer GPUs, and the numbers behind the claim are real enough to check. Zagitta/gpu-unlock targets RTX 30-series cards (GA102, GA103, GA104, GA106 chips, with results reported on the 3090 Ti, 3090, 3080, 3070, and 3060 12GB) and has picked up 8,200 stars and 859 forks on GitHub. JustAResearcher/gpu-unlock-ada does the same for Ada-generation cards, mainly the RTX 4090, with experimental support for the 4090 D, 4080, 4070 Ti SUPER, 4070, and 4060 Ti, and sits at 3,400 stars and 459 forks.

Both tools work the same way: a kernel module loads before Nvidia’s own driver binds to the card, writes new values to the GPU’s speed-selector registers, and forces a function-level reset, all without touching the stock driver or firmware files Nvidia ships. On a 3090 Ti, the author measured BF16 dense GEMM throughput going from 87.5 to 157.4 TFLOPS, a jump that lines up with the long-running theory that Nvidia deliberately caps matrix-math instruction issue rates on gaming silicon to protect the pricing gap with its data-center cards, a claim Nvidia has never confirmed or denied. On the RTX 4090, instruction-level FP16 throughput went from 165.6 to 331.6 TF/s in the author’s own testing.

Real-world gains land well short of that near-2x headline number once you’re running an actual workload instead of a synthetic benchmark. GEMM tests on the 4090 showed 12% to 34% gains for FP16 and 16% to 20% for BF16 depending on matrix shape, with some shapes unchanged entirely. LLM inference barely moves either: prefill improved 14% to 36% in testing, while decode and INT8 performance stayed flat, because token generation lives and dies on memory bandwidth rather than raw matrix throughput, and neither tool touches memory speed. If you’re fine-tuning LoRAs or training from scratch, the gains are real and worth having. If you’re just running a chatbot locally, this unlock won’t do much for you.

Both projects are upfront that this isn’t a plug-and-play install. The Ada tool requires a full cold boot for the first unlock attempt, warns against running nvidia-smi or CUDA tools beforehand since they can bind the driver first, and says a kernel upgrade means regenerating and re-validating the initramfs image. Suspend resets the speed-selector values on both cards, so the module has to re-unlock on resume, and hibernation isn’t supported at all. Neither README reports bricked hardware, but both are explicit that most of their performance data comes from a small number of individually tested cards, not a broad validation pass.