AI Briefinggpu-unlock: RTX 3090 BF16 compute speed reaches 157 TFLOPS
Zagitta/gpu-unlock
About the project
A Linux kernel module that raises the FP16 and BF16 instruction issue rate of the Nvidia GeForce 3000 series to professional-grade levels. FMA compute speeds improve by up to 2x on RTX 3090 Ti, 3080, 3070, and 3060 equipped with GA102, GA103, GA104, and GA106 chipsets.
Unlike UEFI patches or NVIDIA driver modifications, it performs an initialization step before the driver binds to the GPU. This eliminates the need to manually apply patches every time NVIDIA releases a new driver version. It runs automatically from initramfs at boot and supports display output cards as well as cards passed through to VMs.
For the 3090 Ti, BF16 GEMM performance increases from 87.5 TFLOPS to 157.4 TFLOPS, and TF32 from 42.6 TFLOPS to 82.6 TFLOPS. It is particularly effective for LLM prefill stages, diffusion model inference, and FlashAttention operations. LLM decoding and INT8 operations, which are bottlenecked by memory bandwidth, see almost no performance change.
The license is GPL-2.0-only, and the generated firmware is derived from CMP reference code. User reports indicate that output values are bit-identical before and after the change. Performance gains may be reduced in environments with low power limits, so benchmarks from environments with a power limit of 450W or higher should be referenced.
Zagitta/gpu-unlock
Your GeForce runs AI at half speed. This fixes it for 20/30/40 series cards. Up to 2× FP16 / BF16 / TF32 tensor throughput
C
This introduction was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report errors, attribution issues, or removal requests via Contact.