USB UEFI tool lifts Nvidia GeForce FP16 tensor output up to 91%

A developer's laptop RTX 4060 gained 65% more FP16 throughput — from 29.89 to 49.33 TFLOPS — after an open-source UEFI tool removed Nvidia's half-speed tensor limiter. An owner-reported desktop RTX 4080 did better still, rising from 100.37 to 192.18 TFLOPS in BF16 matrix multiply, a 1.91x gain.

The tool, called the UEFI pre-driver NVIDIA GPU unlocker or nv-uefi-unlocker, boots from a FAT32 USB stick before Windows or Linux loads. It removes the tensor throttle Nvidia applies to GeForce cards, which sell below the professional Quadro and RTX A-series parts that carry full-rate tensor units. The project runs entirely at the firmware stage, so users need no VBIOS flash and no patched driver.

How does the unlock work?

The tool applies the community V67 payload with read-back verification. That payload tricks Nvidia's signed Booter firmware into opening the write lock on two speed-selector registers, allowing the tensor units to run at full rate. Read-back verification makes the operation safer than earlier experimental builds, which lacked confirmation that writes had landed.

Previous community unlocks worked mainly through manually installed Linux modules, with Windows support experimental. The new release covers both operating systems from one USB boot. The download ships only scripts and self-built binaries, with no Nvidia firmware included. One script fetches Nvidia's public 610.43.03 driver packages, verifies them, and rebuilds the firmware files locally on the user's machine.

The project credits Zagitta's gpu-unlock as its porting source. It also builds on the RTX 40 fork gpu-unlock-ada and an earlier Windows EFI tool.

What does the user have to do?

The workflow is short but repeats at every boot.

• Download the release and run the build script.
• Copy the rebuilt files and a BOOTX64.EFI to a FAT32 USB stick.
• Boot from the stick via the one-time boot menu and confirm with the spacebar within 30 seconds.
• Control passes to the normal OS loader afterward.

The system must boot in UEFI mode with Secure Boot off, or set to trust a self-signed EFI file. For game anti-cheat systems that require Secure Boot, the tool "CANNOT bypass this requirement," a developer team member wrote on Reddit. RTX 20 owners must run a bundled script once to switch the Windows driver into GSP firmware mode, then restart. On Linux, runtime power management can power down an idle GPU and silently erase the unlock; the FAQ documents workarounds.

Which cards are confirmed working?

Compatibility is routed by chip family, and coverage does not imply per-card testing.

• RTX 20 (Turing): "fully compatible; some models have known issues." Confirmed on Windows: RTX 2070, RTX 2060 Super. The RTX 2080, 2080 Ti and 2080 Super fail at boot or handoff.
• RTX 30 (Ampere): "fully compatible." Confirmed: RTX 3080 including LHR variants, 3060 Ti, 3060 12GB, 3080 Ti Laptop, 3070 Laptop.
• RTX 40 (Ada Lovelace): "fully compatible on laptops and desktops." Confirmed: RTX 4090, 4080, 4070 Ti Super, 4070 Super, 4070, 4060 Ti 8GB and 16GB, 4060 Laptop, 4050 Laptop.
• GTX 16: "theoretically compatible," verification pending. Pre-Turing cards, GA100, and RTX 50-series are unsupported.

What gains do owners report?

The project logs before-and-after results on the same test per card, but the tests differ across rows and are not comparable card to card.

Card	Test	Stock (TFLOPS)	Unlocked (TFLOPS)	Gain
RTX 2070	FP16 matrix multiply	36	60–63	+67–75%
RTX 3060 12GB	FP16 matrix multiply, 4096	25.37	45.79	1.80x
RTX 3070 Laptop	FP16 matrix multiply, 8192	35.43	56.04	1.58x
RTX 4060 Laptop	FP16 matrix multiply	29.89	49.33	1.65x
RTX 4080	BF16 matrix multiply, 4096	100.37	192.18	1.91x
RTX 4090	FP16 matrix multiply, 4096	118.42	171.52	1.45% power-limited

The developer's own desktop RTX 2070 moved from roughly 34 TFLOPS to about 62 TFLOPS in FP16 matrix multiply at 8192.

What are the caveats?

Gains depend on each card's power and thermal limits. The RTX 4090 result, at 1.45x, was explicitly power-limited. The project says the unlock can combine with the nvoc power-limit unlock but warns that doing so "can exceed factory power and thermal envelopes." A per-card ROM hash check, required by the project's own safety rules, is not built yet. Persistence on Thunderbolt eGPUs is "not guaranteed." Zagitta, whose original gpu-unlock started the GeForce unlocking effort, states in its FAQ: "I don't run windows and don't plan to maintain it," and points Windows users to a separate port while warning against unofficial Windows installers.

Laptops need the discrete GPU powered at entry, so dGPU-only or MUX mode is recommended. Laptop results are logged card by card.

The RTX 2080 and RTX 2080 Ti remain under review with no dates promised. A boot-from-disk toggle is in the works to simplify the process further. RTX 50-series owners must wait: the team is still working to crack Blackwell cards, which carry "a lot more security" than it anticipated.