UEFI-level pre-driver Nvidia GPU unlocker now available for RTX 2xxx, 3xxx and 4xxx GPUs on Windows and Linux

A few days ago, GPU modder Zagitta released a GPU unlock utility that essentially doubles the compute performance of the Tensor cores on Nvidia’s RTX 2000 / 3000 / 4000 series GPUs mostly in FP16 and FP32 AI workloads. This implementation only supports Linux distros, however, but other modders immediately started to work on a Windows version. Today, Chinese modder physicsdolphin released what seems to be the first unlock utility supporting all RTX 2000 through 4000 series GPUs on Windows, as well as Linux.

Physicsdolphin’s github repository notes that the utility is a UEFI pre-driver NVIDIA GPU unlocker coded in Rust. In the UEFI DXE window before the NVIDIA driver loads, it performs the complete firmware transaction: SEC2 scrubber → FWSEC-FRTS (WPR2) → signed Booter booting GSP-RM → (gated) V67 four-write payload → read-back verification → FLR → chained handoff back to the Windows Boot Manager. The unlock is a register override applied in the pre-driver window and replayed from a USB stick on every boot. Unfortunately, this means that any anti-cheat system that needs Secure Boot to be enabled will not work with this utility.

As of the latest 2.1.6 version, the entire RTX 3000 series has been tested and proven fully compatible (GA10x device-ID ranges 0x22xx/0x24xx/0x25xx are fully routed). RTX 2000 and 4000 series are still waiting for full compatibility confirmation from user-submitted reports on github. For the RTX 2000 series in particular, users are required to run a GSP firmware-mode script and restart the system. This procedure is not needed for the other two series.

User-submitted reports confirm that FP16 and FP32 workloads are seeing the most gains (usually between 50-80%), but BF16 workloads can also see notable boosts. One RTX 4070 Ti Super user managed to somehow boost FP 16 and BF16 workloads by over 2x: FP16 4096 57.00 → 150.93 (2.65×), BF16 4096 57.51 → 127.32 (2.21×), FP16 8192 55.50 → 127.28 (2.92×), BF16 8192 54.97 → 128.39 (2.34×).

Support for the GTX 16xx series that retains the dedicated integer core for FP operations is also under development, with gains similar to the RTX 2000 series.