Open-source nv-uefi-unlocker tool removes Nvidia tensor limiter on RTX GPUs via bootable USB

The nv-uefi-unlocker project has released an open-source tool that boots from USB to remove Nvidia's artificial tensor performance limiter on consumer GeForce GPUs. Running before the OS loads, the UEFI payload tricks Nvidia's signed Booter firmware into unlocking write-protected speed-selector registers, enabling full-rate FP16 tensor operations on RTX 20, 30, and 40-series desktop and laptop cards. A developer's laptop RTX 4060 demonstrated 65% higher FP16 math performance after unlocking. The tool eliminates previous requirements for manual Linux kernel module installation or experimental Windows support, making the hack accessible to a broader audience through a simple FAT32 USB boot process with 30-second user confirmation.

The release contains only scripts and self-built binaries, with no redistributed Nvidia firmware. A bundled script downloads and verifies Nvidia's public 610.43.03 driver packages, then rebuilds firmware files locally. The project incorporates community work including Zagitta's gpu-unlock, the gpu-unlock-ada RTX 40 fork, and an earlier Windows EFI tool, adding read-back verification for safety. The unlock is non-persistent across power cycles, requiring re-execution at each boot. Systems must use UEFI mode with Secure Boot disabled or configured to trust a self-signed EFI, which conflicts with some anti-cheat requirements. RTX 20-series cards need a one-time GSP firmware mode switch on Windows, while laptop users should run in dGPU-only or MUX mode.

How the UEFI unlocker works and its constraints

• UEFI pre-driver execution: The tool runs as BOOTX64.EFI from a FAT32 USB stick during firmware initialization, before the OS Nvidia driver loads, avoiding any need to flash VBIOS or patch drivers.
• Register unlock mechanism: Uses the community V67 payload with read-back verification to manipulate Nvidia's signed Booter firmware, unlocking write access to two speed-selector registers that govern tensor unit clock rates.
• Local firmware rebuild: Ships no Nvidia proprietary firmware; instead downloads public 610.43.03 driver packages, verifies them cryptographically, and rebuilds required firmware files on the user's machine.
• Non-persistent state: The unlock does not survive power loss, requiring execution at every boot; on Linux distributions with runtime power management, GPU idle power-down can silently erase the unlock, requiring documented workarounds.
• Platform requirements: Demands UEFI boot with Secure Boot off or self-signed EFI trust; RTX 20-series requires GSP firmware mode on Windows; laptops need dGPU-powered entry state with MUX or dGPU-only mode recommended.

Why it matters

This tool significantly broadens access to consumer GPU AI compute capabilities that Nvidia artificially segments, potentially accelerating local AI inference and training workflows for researchers and enthusiasts who cannot afford datacenter cards. For the Linux and open-source ecosystem, it represents a notable reverse-engineering achievement that operates entirely through open-source UEFI tooling without proprietary redistribution, though users must weigh the security trade-offs of disabling Secure Boot and the maintenance burden of per-boot execution.