gpu-unlock on GitHub: changing FP/BF16 issue rate on Nvidia RTX 30

⚡ The gist in 5 seconds - A repository, Zagitta/gpu-unlock, has appeared on GitHub — it's about the FP/BF16 instruction issue rate on Nvidia RTX 30 (Ampere) GPUs. - The topic made it to Hacker News: 143 points and 53 comments. - There's little concrete detail behind the story: no list of cards, no driver versions, no performance measurements — everything is in the repository itself. ### 🔍 What was found The story centers on the Zagitta/gpu-unlock repository, which surfaced on Hacker News with the description "FP/BF16 instruction issue rate on Nvidia 3000's cards." It concerns Nvidia RTX 30-generation GPUs built on the Ampere architecture and the rate at which the chip issues FP and BF16 format instructions — that is, throughput at the SM scheduler level, not frequency overclocking. This is not an official Nvidia release: the project lives on GitHub, and the company has nothing to do with it. What the code actually does doesn't follow from the story — there are no version numbers, no list of supported models, no build flags, and no before/after measurements in the description. All of that has to be dug out of the README and source code. The only thing known for sure is the subject: FP/BF16 and consumer RTX 30 cards, where half the lineup has historically been cut back on a number of compute parameters compared to professional accelerators. ### 💡 Why it matters RTX 30 is far from a flagship these days, but it's exactly these cards that are massively used in home builds for local inference — and FP16/BF16 is one of the key formats there. Any attempt to raise the issue rate promises a free performance boost in a situation where the usual advice is simply to buy new hardware. But don't celebrate too early: there's no official support, no one has guaranteed stability or computational correctness, and the figures in the Hacker News comments are opinions, not benchmarks. Test on your own card at your own risk.