A new open-source tool lets owners of GeForce RTX 20, 30 and 40-series cards switch off the cap that holds some of Nvidia’s AI maths at half speed. It runs from a USB stick before Windows starts, so nothing on the card or in the driver is permanently changed.
The cap affects tensor-core work in FP16 and similar formats, which matters for local AI models, image generation and some video tools. It does not make games run faster. The project is a community effort, not an Nvidia feature, and the card goes back to its normal state when the power is cut.
How the USB unlocker works
The project is called nv-uefi-unlocker and is released under the GPL-2.0 licence. You build a FAT32 USB stick from its release files, boot the PC from it using the one-time boot menu, and press Space or Enter on the version screen within 30 seconds. After that, control passes to the normal Windows loader.
Because this happens before the Nvidia driver loads, there is no VBIOS flash and no patched driver. The catch is that the unlock disappears at power-off, so you repeat it after every cold boot. The PC must boot in UEFI mode with Secure Boot off, or with the tool’s self-signed file trusted. The project’s team says it cannot get around a game’s anti-cheat that demands Secure Boot.
The release contains no Nvidia firmware. A script downloads Nvidia’s public driver packages and rebuilds the files it needs on your own machine, which means the first run needs an internet connection.
Which cards, and how much faster
The README lists RTX 30 and RTX 40 cards, desktop and laptop, as fully compatible. RTX 20 cards are marked mostly compatible with known issues on some models, and need a bundled script run once to put the Windows driver into GSP firmware mode. The GTX 16 series is only “theoretically” supported. Pascal and older cards, the data-centre GA100 and the entire RTX 50 series are not supported, and the developers say the 50-series carries much more security than they expected.
The README claims up to about 2.6 times the FP16, BF16 and TF32 tensor throughput and up to 1.9 times in FP8, with the real gain capped by power limits. In the numbers Tom’s Hardware reported, the developer’s own RTX 4060 laptop gained 65% in FP16 maths, and a desktop RTX 2070 rose from roughly 34 to 62 TFLOPS in an FP16 matrix-multiply test. Those are the developer’s results, and the project warns that its tests differ from card to card, so they should not be compared across rows.
RTX 2080, 2080 Ti and 2080 Super cards currently fail at boot or at the handoff, according to the same report.
The risks are yours
Combining the unlock with the project’s power-limit tool can push a card beyond its factory power and heat limits, and the README says cooling and power headroom are the owner’s responsibility. If the tool hangs or shows an error, the instructions are to power off fully, pull the USB stick and follow the recovery notes rather than retry in the same session. A per-card ROM hash check that the project says it needs is not built yet.
The developers say a boot-from-disk option is planned to make the process easier. No dates are given for the RTX 2080 family or for RTX 50 cards.
