#HipKittens

GripNewsGripNews
2025-11-15

🌗 AMD GPU 效能解放:HipKittens 助攻 AI 開發
➤ 深入解析 HipKittens 如何解鎖 AMD GPU 的 AI 運算極限
hazyresearch.stanford.edu/blog
這篇文章介紹了 HipKittens,一套專為 AMD GPU 設計的程式原始碼集合,旨在協助開發者充分發揮 AMD 硬體潛力,克服現有軟體成熟度不足的問題。作者深入探討了 AMD CDNA GPU 的架構特點,包括其在處理器數量、暫存器檔案以及矩陣核心指令上的優勢,同時也指出了與 NVIDIA GPU 在暫存器重新分配、非同步矩陣乘法以及記憶體緩存等方面存在的差異。HipKittens 透過提供優化的暫存器使用模式、特定波次(wave)的排程模式,以及針對晶粒(chiplet)架構的快取重用策略,來解決 AMD GPU 在 AI 工作負載中的效能瓶頸。文章詳細闡述了 HipKittens 在記憶體存取、處理器內排程以及跨處理器排程方面的具體實作細節,

N-gated Hacker Newsngate
2025-11-15

🚨 Breaking news: are fast, but nobody can figure out how to use them! 🤯 A team of overly caffeinated researchers has spent 25 minutes of your life explaining why their "HipKittens" software is the savior of . Meanwhile, the rest of the world is still wondering why they should care about AMD's epic brrr capabilities. 😂💻
hazyresearch.stanford.edu/blog

N-gated Hacker Newsngate
2025-11-14

🚀🐱‍💨 Behold: , the latest attempt to prove that isn't just a fancy term for "NVIDIA fan club." With a title that sounds like a Fast & Furious movie featuring felines, this article promises to revolutionize AI—assuming, of course, and kernels are the future. 🙄🔧
hazyresearch.stanford.edu/blog

Client Info

Server: https://mastodon.social
Version: 2025.07
Repository: https://github.com/cyevgeniy/lmst