Zagitta Releases gpu-unlock, Lifts FP16/BF16 Caps on RTX 3000s
TL;DR
- Zagitta's gpu-unlock 1.0.0 loads via initramfs before Nvidia's driver binds, lifting FP/BF16 instruction issue caps on GA102 through GA106 consumer cards.
- On an RTX 3090 Ti the project reports FP16 climbing to 118.39 TFLOPS from 87.46 baseline, with BF16 reaching 157.41 TFLOPS from 87.50.
- Reported GEMM gains span +25% to +81% depending on card model and power configuration; the kernel module ships GPL-2.0-only.
Zagitta has posted gpu-unlock 1.0.0, a Linux kernel module that lifts the firmware cap on FP16 and BF16 instruction issue rates for Nvidia's GeForce RTX 3000-series cards, aiming to put consumer silicon on par with its pro-grade equivalents on matrix workloads. On an RTX 3090 Ti, the project's own benchmarks reportedly climb to 118.39 TFLOPS FP16 against an 87.46 baseline and 157.41 TFLOPS BF16 against 87.50, with TF32 rising to 82.61 from 42.58.
The module targets GA102, GA103, GA104 and GA106 chips, covering the 3090 Ti, 3090, 3080 Ti, 3080, 3070 Ti, 3070 and 3060 12GB, and loads during early boot through initramfs before the Nvidia driver binds. Per the project README on GitHub, the project implements 'the CMP unlocker's V67/PLM technique with pre-driver timing' to run firmware unlock sequences and write speed selector values without patching the proprietary driver.
Zagitta reports GEMM gains between +25% and +81% depending on card model and power configuration. The kernel module and local packaging ship under GPL-2.0-only, while generated firmware derives from pinned references and the Nvidia firmware retains its original proprietary terms. Install is a `make install` plus appending `gpu_unlock.targets=all` to kernel command line parameters. The release joins a steady run of consumer-hardware hacks surfacing across our open-source feed.
Originally reported by github.com
Read the original article →Original headline: Linux Kernel Module 'gpu-unlock' Enables Pro FP/BF16 Throughput on Nvidia RTX 3000 GPUs