New UEFI tool unlocks extra Tensor throughput on RTX 20 to 40 cards in Windows
Modder physicsdolphin's nv-uefi-unlocker boots from USB before the driver loads, needs no flashing and resets after a full power-off. RTX 50 is unsupported.
Chinese modder physicsdolphin has released nv-uefi-unlocker, a free tool that lifts NVIDIA's Tensor core throughput cap on GeForce RTX 20, 30 and 40 series cards under Windows. It arrived in late October 2026, about a week after modder Zagitta showed the same limit could be removed on Linux, and its project page claims FP16, BF16 and TF32 Tensor throughput gains of up to about 2.6 times.
How it works
Instead of patching NVIDIA's driver or flashing the card's firmware, the unlocker is a UEFI application written in Rust. You build a FAT32 USB stick with the project's script, boot from it, and the tool writes its changes to the GPU before handing control back to Windows, Linux or any other EFI-bootable OS. Nothing is permanent: a full power-off restores stock behaviour, so the unlock has to be replayed on every cold boot. Secure Boot must be off or set to trust a self-signed EFI file, and RTX 20 owners have to enable GSP firmware mode with a PowerShell script first.
The numbers so far
Results posted with the project are big on paper. An RTX 4060 Laptop GPU went from 29.89 to 49.33 teraflops in an FP16 matrix test, and an RTX 2070 climbed from about 36 to roughly 60 to 63 teraflops in FP32-accumulation matrix multiplication. The developer also lists FP8 gains of up to about 1.9 times, image and video generation speedups of 14 to 75%, and DirectML inference operators running up to 94% faster.
What it means for gamers
This is an AI compute unlock, not a frame-rate boost. Wccftech found no convincing evidence that it improves gaming performance or DLSS processing times, and real-world gains depend on power limits, memory bandwidth and the workload. GTX 16 support is experimental and RTX 50 cards are not supported at all. For owners of an older RTX card who run Stable Diffusion-style image generation or local language models, it is a meaningful free upgrade; for pure gaming rigs, there is little reason to touch firmware-level tools.
Risks and next steps
The project is still moving fast, with a run of releases up to v2.1.9-rc1 fixing hangs on specific motherboards and Turing edge cases. Its own instructions say not to retry in the same session after an error or black screen, and to power off fully before booting normally. NVIDIA has not commented.
Source: TechPowerUp; physicsdolphin (nv-uefi-unlocker on GitHub); Wccftech.
