Unlocking AI potential on Nvidia cards

Users can now bypass restrictions on using Nvidia's GeForce RTX graphics cards for AI tasks with a new open-source tool called nv-uefi-unlocker. This tool allows the cards to perform AI calculations without affecting their gaming performance, which remains unchanged.

How the tool works

The tool operates during the UEFI phase, which is the early stage of system boot before Windows and Nvidia drivers load. It injects a payload that modifies specific registers to adjust the speed of Tensor cores, which are crucial for AI computations. These changes are temporary and can be reset by simply cutting the power.

Development and compatibility

The project is built in Rust and is open for public review, crediting earlier efforts like gpu-unlock by Zagitta and gpu-unlock-ada. It has been tested on seventeen models across the RTX 20, 30, and 40 series, though none from the RTX 50 series are supported. Compatibility lists are based on chip families, not all tested models.

Technical requirements and limitations

To use the tool, users must boot in UEFI mode with Secure Boot disabled or configured to accept self-signed files. Some games with anti-cheat systems that require Secure Boot may not launch, and the tool cannot resolve this issue. Additionally, the gain in gaming performance is zero, as the changes only affect AI calculations.

Performance improvements in AI tasks

In controlled laboratory tests, the unlocked cards showed performance boosts ranging from +45% to +91% in AI tasks. For example, an RTX 4060 laptop GPU saw a jump from 29.89 to 49.33 TFLOPS. However, these results vary depending on the card and test conditions, and they are not comparable across different models.

Linux users report limited gains

Linux users of RTX 30 cards have reported smaller performance gains, between 14% and 36%, when running language models. However, text generation performance is nearly unchanged, as it is limited by the speed of the video memory rather than the Tensor cores.