USB Tool Unlocks More AI Performance on Older Nvidia RTX GPUs, but There Are Catches
Your older GeForce RTX card might have more tensor performance available than its current configuration allows. An open-source project called nv-uefi-unlocker now offers a USB-based way to remove Nvidia’s tensor throughput restriction on supported RTX 20, 30 and 40-series GPUs under Windows and Linux.
Tom’s Hardware reports that the developer recorded 65% higher FP16 maths performance on a laptop RTX 4060. That is an eye-catching result, but the distinction matters: this is a tensor-compute unlock, not evidence that your games will suddenly run 65% faster.
What the unlocker actually changes
The utility operates during UEFI startup, before the operating system loads Nvidia’s graphics driver. Using the community’s V67 payload, it gets Nvidia’s signed Booter firmware to permit writes to two registers controlling tensor speed, then checks that the changes took effect.
Unlike a graphics-card firmware flash, this approach does not require replacing the VBIOS or modifying the Nvidia driver. Earlier routes largely involved manually installed Linux modules, while Windows support was experimental.
The download also does not bundle Nvidia firmware. Instead, a supplied script retrieves Nvidia’s public 610.43.03 driver packages, verifies them and reconstructs the required firmware files locally.
Easier to try does not mean hassle-free
Preparation still involves running that script and placing the resulting files, together with a BOOTX64.EFI boot file, on a FAT32 USB drive. Users launch it through a one-time boot menu and confirm with the spacebar within 30 seconds. The unlocker then passes control to the operating system loader.
You need to repeat the process at every boot. The change does not survive power loss.
The PC must also use UEFI boot, with Secure Boot either disabled or configured to trust a self-signed EFI. Crucially, the project cannot bypass a game’s requirement for Secure Boot.
RTX 20-series users face another preliminary step: running a bundled script to enable GSP firmware mode for the Windows Nvidia driver, then restarting.
Check your exact GPU, not just its generation
The project lists RTX 30-series support as fully compatible, alongside desktop and laptop RTX 40-series support. RTX 20-series support comes with known model-specific problems: the RTX 2080, 2080 Ti and 2080 Super are reported to fail during boot or the handoff to the operating system.
GTX 16-series compatibility remains theoretical and unverified. RTX 50-series cards, pre-Turing GPUs and GA100 are unsupported. Support by chip family also does not mean every individual card has been tested.
Laptop owners need the discrete Nvidia GPU powered when the unlocker starts, so dGPU-only or MUX mode is recommended. On Linux, runtime power management can switch off an idle GPU and remove the unlock. Persistence with Thunderbolt external GPUs is not guaranteed either.
What Malaysia and SEA readers should take from this
Our take: if you already own a supported RTX GPU and use it for local AI or FP16 compute work, this is worth watching. The potential benefit is getting more from existing hardware—not a confirmed gaming upgrade or a reason to buy an untested card.
The developer’s desktop RTX 2070 also reportedly climbed from roughly 34 to 62 TFLOPS in an FP16 matrix-multiplication test. However, the published results use different tests across cards, so they are not a reliable cross-GPU ranking.
There are unfinished safeguards too. A per-card ROM hash check required by the project’s own safety rules has not yet been implemented. Power and cooling limits still affect results, and combining this with the separate nvoc power-limit unlock can push hardware beyond factory power and thermal boundaries.
Promising enthusiast project? Definitely. A no-caveats upgrade for every RTX owner? Not yet.
