Firmware and hardwareWorksSource available
Target: Nvidia 30 series GPUs

GPU-unlock: FP/BF16 instruction issue rate on Nvidia 3000's cards

by Zagitta Repository0 views
Platform
Linux
Outcome
Works

Evidence at a glance

Loading source evidence…

What was done

A Linux kernel module that unlocks FP/BF16 instruction issue rates on consumer Nvidia 30 series GPUs by applying firmware-level speed selector modifications before the proprietary driver binds the hardware.

The GPU-unlock project provides a Linux kernel module designed to increase the FP/BF16 instruction issue rate on consumer Nvidia 30 series graphics cards to match their professional counterparts. The author reports that the module functions by combining the cmpunlocker's V67/PLM technique with a pre-driver timing approach. By loading the module via the initramfs before the official Nvidia driver binds the card, the software writes specific speed selectors (SS0 and SS1) to the hardware. The README documents that this process effectively doubles the instruction issuing speed of FP/BF16 FMA operations on supported architectures, including GA102, GA103, GA104, and GA106. The project relies on firmware extracted from official Nvidia drivers during the build process, though the distributed packages themselves do not contain proprietary archives. Commit logs indicate that AI models, specifically Claude Sonnet 5.5 and Claude Opus 5.5, have contributed to the project by authoring documentation, creating issue templates, and implementing CI workflows. The author notes that while the unlock is intended for Linux, ports for Windows exist, and support for other GPU architectures is under active development.

Reproduce it

setup
Clone the repository, run `make install` to detect the distribution and build the module, then append `gpu_unlock.targets=all` to the kernel command line and reboot.