HN
Today

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

This project unveils a clever Linux kernel module that unlocks the full FP/BF16 performance of Nvidia's 3000 series GPUs, essentially upgrading consumer cards to professional-grade compute power. By manipulating boot-time driver loading, it bypasses Nvidia's artificial segmentation, offering a substantial TFLOPS boost. It's an instant hit on Hacker News for its technical ingenuity, performance optimization, and defiance of hardware limitations.

16
Score
1
Comments
#20
Highest Rank
2h
on Front Page
First Seen
Oct 24, 2:00 PM
Last Seen
Oct 24, 3:00 PM
Rank Over Time
2620

The Lowdown

The gpu-unlock project introduces a highly technical solution for extracting more compute performance from Nvidia's consumer-grade 3000 series GPUs. It specifically targets the FP/BF16 instruction issue rate, bringing cards like the GA102, GA103, GA104, and GA106 closer to the capabilities of their professional counterparts.

  • Core Functionality: The tool is a Linux kernel module that operates early in the boot process, leveraging a V67/PLM technique combined with pre-driver timing. It writes specific speed selectors (SS0 and SS1) and performs an FLR before the standard NVIDIA driver takes over.
  • Performance Impact: The primary effect is a doubling of the instruction issuing speed for FP/BF16 FMA operations into FP32. For instance, a 3090 Ti saw BF16 performance jump from 87.50 TFLOPS to 157.41 TFLOPS in dense GEMM benchmarks.
  • Installation & Usage: Installation involves cloning the GitHub repository, running make install, and appending gpu_unlock.targets=all (or specific PCI IDs) to the kernel command line to activate the module.
  • Boot & Resume Integration: The module integrates into the initramfs, ensuring it loads before the official NVIDIA driver binds to the GPU. It also handles re-unlocking cards after a suspend-to-RAM event to maintain performance.
  • Licensing: The kernel module and local packaging are licensed under GPL-2.0-only, while NVIDIA firmware components retain their original terms.

This project provides a significant, software-driven performance upgrade for owners of eligible Nvidia GPUs, making high-end AI and scientific computing more accessible without requiring hardware modifications.