gpu-unlock repo claims to unlock hidden tensor core speed on RTX 3090 Ti: 87 to 157 TFLOPS
jfischoff · x · 2026-10-25
A viral greentext-style post claims a new GitHub repo, gpu-unlock, flips speed selector bits before the driver loads on Ampere cards, unlocking full-rate FP16/BF16 (FP32 accumulate) matmul on the RTX 3090 Ti — benchmarks cited show a jump from 87 to 157 TFLOPS, effectively a free generational leap for a 6-year-old GPU. The claims are unverified at time of posting, and the post jokes about eBay listings repricing used 3090s as "BF16 unlocked." If confirmed, it would be a big win for local LLM hobbyists still running 24GB Ampere cards; treat with caution pending independent reproduction.
More from Fun
- Isaacson: Musk's first two companies made him ~$100M each — he bet it all on Mars — elonmusk · 2026-10-25
- "The city will be down one bus, not one CEO" — Sam Altman interview meme goes around — granawkins · 2026-10-25
- Google's AI Overview politely plays customer service after a user hurls German insults into Search — Glass-League-8259 · 2026-10-25
- One chart maps all of new media by 'requires braincells' vs bias — LessWrong lands top-right — FinanceYF5 · 2026-10-25
- Activation steering recipe turns Qwen 3.5 4B into "evil qwen" by nudging two neurons — rachittshah · 2026-10-25
- After 24 hours of testing, this dev's hill to die on: Kimi 3.1 is near AstrA — bookwormengr · 2026-10-25