Open-source USB tool unlocks Nvidia GPU AI performance on Windows

Bypassing the hardware limit

A new open-source project called the nv-uefi-unlocker provides a method for users to bypass Nvidia's hardware-level AI performance limits. This specific limit affects the tensor units, which are the parts of the GPU designed to handle the heavy math required for artificial intelligence. By removing this throttle, users can significantly increase the speed of AI-related tasks on their existing hardware.

The tool targets several popular generations of graphics cards, including the RTX 20, 30, and 40-series. It works on both desktop and laptop configurations, making it accessible to a wide range of PC owners. While the project is still developing, it aims to simplify a process that was previously much more difficult for the average user to execute.

As Tom's Hardware reported, the tool can lift FP16 math performance by as much as 65% on certain models. For example, the developer noted that an RTX 4060 laptop GPU saw a significant jump in performance after applying the unlock. This provides a substantial boost for anyone running local AI models or other compute-heavy applications.

How the UEFI unlock works

The nv-uefi-unlocker operates in a unique way by running from a bootable USB stick before the Windows or Linux operating system even starts. Because it runs during the UEFI stage, it can interact with the hardware before the official Nvidia drivers are loaded. This means users do not have to flash their card's video BIOS or modify the system drivers to achieve the performance boost.

Technically, the tool uses a community-developed payload to trick the signed Booter firmware into opening a write lock on two specific speed-selector registers. Once these registers are unlocked, the tensor units are allowed to run at their full potential rather than being held back by the manufacturer. The project includes a read-back verification step to ensure the unlock was successful and to make the process safer than previous experimental versions.

To use the tool, a user must format a USB stick to FAT32 and copy the necessary scripts and binaries onto it. The system must be set to UEFI mode, and Secure Boot must be disabled or configured to trust a self-signed EFI. Because the unlock happens at the firmware level, it does not persist through a power loss, meaning the user must boot from the USB and run the tool every time they want the extra performance.

Compatibility and requirements

The tool is designed to be broadly compatible across the RTX 20, 30, and 40-series families. While most cards are listed as fully compatible, some specific models within the RTX 20-series may have known issues. The project routes its support by chip family, so a specific card being listed does not necessarily mean it has been individually tested by the developers.

Laptop users have specific requirements to ensure the unlock works correctly. The discrete Nvidia GPU must be powered at entry, which means the user should ideally use a dGPU-only or MUX mode. Additionally, on Linux systems like Ubuntu, the user may need to address runtime power management, which can sometimes power down the GPU and erase the unlock when the card is idle.

Certain hardware remains unsupported by the current release. The RTX 50-series cards are not supported because they feature significantly more security than the team anticipated. Furthermore, the tool cannot bypass the Secure Boot requirements for games that use anti-cheat software. Other cards, such as the GA100 and pre-Turing models, are also currently excluded from the project.

Performance and safety risks

The performance gains from this unlock are highly dependent on the specific hardware and its power limits. The developers warn that the unlock can be combined with other power-limit tools, but doing so might exceed the factory power and thermal envelopes of the card. Users should be aware that pushing the hardware beyond these limits can lead to overheating or other stability issues.

Safety is a primary concern for the project, which is why it includes a verification step to confirm the registers were successfully modified. The release does not include any Nvidia firmware in the download; instead, it uses a script to grab public driver packages and rebuilds the necessary files locally. This approach is intended to make the tool less intimidating and safer for users to try compared to older methods.

The project is currently being maintained by a community of developers who are building on several previous projects. One of the original creators of the underlying unlock method has stated they do not plan to maintain the Windows port. Users are encouraged to be careful with unofficial installers and to follow the guidelines provided in the project's FAQ to ensure a stable experience.

What comes next for the tool

The development team is actively working on several improvements to make the tool more user-friendly. They are currently working on a boot-from-disk toggle, which would eventually allow users to run the unlock without needing to manually boot from a USB stick every time. This would significantly streamline the process for those who want to use the extra performance regularly.

While the current version is functional, there are still several features that have not been implemented yet. For instance, a per-card ROM hash check is not yet built into the tool, which is one of the project's own safety rules. The developers are also still reviewing the RTX 2080 and RTX 2080 Ti to see if they can be added to the supported list.

For now, the tool remains a manual process that requires some technical knowledge to set up correctly. However, it represents a significant step forward in making high-performance AI math accessible on consumer hardware. As the project progresses, it may become even easier for enthusiasts to squeeze more power out of their Nvidia GPUs for local AI workloads.