Hao AI Lab· @haoailab · X·· 2026-07-14AI 评分48
AI 导读
🚀 注意力-前馈网络分离在最新的机架级 GPU 系统上还能赢吗?我们构建了 FastAFD,一个面向 GB200 NVL72 的开源 AFD 运行时:72 块 Blackwell GPU 位于同一个 NVLink 域内。它将每 GPU 解码吞吐量提升了 1.35-1.45 倍。🧵 代码:https://github.com/hao-ai-lab/FastAFD 博客:https://haoailab.com/blogs/fastafd/
正文 · 原文
🚀 Can Attention-FFN Disaggregation still win on the newest rack-scale GPU systems?
We built FastAFD, an open-source AFD runtime for GB200 NVL72: 72 Blackwell GPUs in one NVLink domain.
It improves per-GPU decode throughput by 1.35-1.45×. 🧵
Code: https://github.com/hao-ai-lab/FastAFD
Blog: https://haoailab.com/blogs/fastafd/
来源:Hao AI Lab · x.com