Sudo su (@sudoingX)Qwen 3.5-0.8B의 24개 레이어를 하나의 fused CUDA 커널로 처리해 CPU 왕복 없이 단일 커널 런치만으로 고성능을 달성했다는 내용이다. RTX 3090이 411 tok/s로 Apple M5 Max의 229 tok/s를 앞질렀다고 소개하며, CUDA 최적화와 추론 성능 개선의 인상을 준다.https://x.com/sudoingX/status/2042093043533410339#cuda #qwen #llm #gpu #optimization
Related
@leicestergazette Why? Because billionaires are funding racists, as their bang bang bang on democracy tool. Paid thugs.L...
@leicestergazette Why? Because billionaires are funding racists, as their bang bang bang on democracy tool. Paid thugs.Like pedos, money to hurt society. By billionaires, they are ...
The 2-Second Rule: How to Make Your Website Feel Like Magic in 2026, by @webdesignerdepot.bsky.social:https://webdesigne...
The 2-Second Rule: How to Make Your Website Feel Like Magic in 2026, by @webdesignerdepot.bsky.social:https://webdesignerdepot.com/the-2-second-rule-how-to-make-your-website-feel-l...
ミッドナイト、簡単なようで難しいですね……値上げ前には届かないけど4万円切り！ Apple Watch SE 3が39,192円／Amazonで値引き https://ascii.jp/elem/000/004/426/4426023/?r...
ミッドナイト、簡単なようで難しいですね……値上げ前には届かないけど4万円切り！ Apple Watch SE 3が39,192円／Amazonで値引き https://ascii.jp/elem/000/004/426/4426023/?rss#Apple #LLM #news #bot