Oriel — AI & technology worth noticing, every day.

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

Repository · https://github.com/zagitta/gpu-unlock

First seen by Oriel on .

AI summary

gpu-unlock is a Linux kernel module that configures selected Nvidia RTX 3000 GPUs before the NVIDIA driver binds, doubling FP/BF16 instruction issue rate on supported GA102, GA103, GA104, and GA106 cards.

  • It combines the CMP unlocker's V67/PLM technique with pre-driver timing, writes speed selectors, and performs a function-level reset before driver binding.
  • On a 3090 Ti, reported 8192³ GEMM throughput rose from 87.50 to 157.41 TFLOPS for BF16 and from 42.58 to 82.61 TFLOPS for TF32.
  • User reports show BF16 GEMM gains varying by card and power limit; an OEM RTX 3080 Ti 20 GB completed the unlock but remained at half rate.
  • Matrix-heavy tasks including LLM prefill, diffusion, and FlashAttention can benefit, while memory-bandwidth-limited LLM decode and INT8 work stay about the same.
  • The module is opt-in through the kernel command line, runs from the initramfs, and can reapply the unlock after suspend to RAM; hibernation is not covered.

#nvidia gpus #linux kernel #gpu compute #bf16

Generated by gpt-5.6-sol. Check the original for details.

Metric trends

Last 30 days · Up to 120 recorded samples per source and metric, with reposts kept separate · First and latest retained

Points

Score reported by the source.

  • Hacker News: 341, latest observation ; +336 over observed period since

Comments

Comment count reported by the source.

  • Hacker News: 80, latest observation ; +79 over observed period since

Aliases

SourceExternal IDURL
Hacker News50219700https://github.com/Zagitta/gpu-unlock Read discussion