Free AI performance boost: New tool unlocks the potential of Nvidia graphics cards

A programmer with the nickname Zagitta has already introduced a tool for unlocking Nvidia graphics cards that significantly increases the performance of Tensor cores for AI tasks, but it only works on Linux. Recently, a programmer with the nickname physicsdolphin issued a solution in the form of UEFI pre-driver, which also brings support for Windows.

The program is written in Rust and runs in a special UEFI DXE environment before the official driver is loaded. In this step, it performs several interventions in the firmware, from cleaning SEC2 and preparing FWSEC to starting GSP-RM and finally writing and verifying commands. Unlocking is done by overwriting the registers from the USB stick at each boot. Since the Secure Boot function must be turned off, some games with anti-cheat protection may not work.

Version 2.1.6 has already been tested and confirmed for the entire RTX 3000 series. User confirmations are still being collected for the RTX 2000 and RTX 4000 series. RTX 2000 card owners must first run the GSP mode script and then reboot the system.

Users are reporting significant speed increases when working with AI models. Improvements are typically between 50 and 80 percent for FP16 and FP32 workloads, and performance gains are also noticeable for BF16. An RTX 4070 Ti Super user more than doubled performance with the tool: from 57 to almost 151 TFLOPS for FP16, and from 57.5 to 127 TFLOPS for BF16. Support for the GTX 16xx series, which feature separate integer cores, is also in development.