What is gpu-unlock?
gpu-unlock is a Linux kernel module for developers and system administrators running AI workloads on supported NVIDIA GeForce GPUs. It changes tensor speed selectors before the unmodified NVIDIA driver binds the card. The intended benefit concerns FP16, BF16 and TF32 matrix operations with FP32 accumulation, rather than general GPU performance.
Deployment uses local builds and native packages for Arch, Debian/Ubuntu and Fedora/RHEL distribution families. The packages place the module and generated firmware in the initramfs. Unlocking requires an explicit kernel command-line opt-in, either for all supported cards or selected PCI devices. The process does not require a patched NVIDIA driver or a VBIOS flash.
The project describes Ampere support as stable, while Ada Lovelace and Turing support remain experimental. Blackwell is unsupported. Application gains depend on tensor computation intensity and available power headroom. Memory-bound token generation, gaming and FP32 CUDA workloads are not expected to improve.
Key capabilities
- Tensor rate adjustment Removes the documented rate cap for supported FP16, BF16 and TF32 operations with FP32 accumulation.
- Pre-driver execution Applies the unlock during early boot before the stock NVIDIA driver binds the GPU.
- Selective targeting Enables explicit opt-in for supported cards through kernel command-line settings and PCI device selection.
- Native packaging Builds and installs packages for supported Arch, Debian/Ubuntu and Fedora/RHEL distribution families.
- Resume handling Reapplies the unlock after suspend to RAM when the card remains eligible; hibernation is not covered.
- Rollback control Disables unlocking when its target option is removed, with package removal also removing the module and initramfs configuration.
Where gpu-unlock fits
gpu-unlock fits beneath existing Linux AI software stacks rather than replacing an inference engine or training framework. Its documented workloads include LLM prompt processing, BF16 fine-tuning and FP16/BF16 image generation. The project also reports limited gains for quantized GGUF workloads in llama.cpp and Ollama, chiefly at longer context lengths.
Before deployment, operators should verify the exact GPU device against the supported list, check distribution and initramfs handling, and benchmark representative workloads under normal power limits. Laptop power management can silently reset the unlock when the dedicated GPU powers off. Boot logs and suspend/resume behavior need validation, and removing the kernel target option provides a way to return to an unmodified boot path.
Install
make installReviews of gpu-unlock
Be the first to share your experience with gpu-unlock.
Write a review
Loading review form…
Alternatives to gpu-unlock
DeepInferDeepInfer is an Apache-licensed DeepSeek inference engine for Linux, NVIDIA SM100 GPUs and Huawei Ascend 950 NPUs, supporting offline inference.
LocalAILocalAI is an MIT-licensed inference server that runs language, speech, image, video and vision models on CPUs or GPUs and serves them through OpenAI, Anthropic, Ollama and ElevenLabs-compatible APIs, with backends pulled on demand.
OllamaOllama is an MIT-licensed tool for downloading and running open models such as Gemma, Qwen, DeepSeek and gpt-oss on your own machine. It ships a desktop app, CLI and local REST API, and offers cloud-hosted models on a pay-as-you-go or subscription basis.