MiniMax (official)· @MiniMax_AI · X·· 2026-07-11AI 评分47
AI 导读
MiniMax 与 Fireworks AI 合作，为 M3 模型在 NVIDIA Blackwell（B200）上推出新内核。该内核采用 KV-stationary 设计，每个选中的 KV 块仅读取一次，解决了长上下文稀疏注意力中数据依赖块选择破坏内存访问速度的问题，在 B200 上达到约 980 TFLOP/s 的算力。
正文 · 原文
Sparse attention only matters if the systems can preserve the theoretical gains.
@FireworksAI_HQ new M3 kernel on Blackwell uses a KV-stationary design to read each selected block once, reaching ~980 TFLOP/s on a B200.
Read the full breakdown below for a closer look at the kernel design and optimizations. 👇
Long-context sparse attention has a catch: data-dependent block selection wrecks memory access kills speed. Our @MiniMax_AI M3 kernel on Blackwell answers it. KV-stationary, each block read once, ~980 TFLOP/s on a B200. See the breakdown here → https://fireworks.ai/blog/kernel-optimization-for-minimax-m3-on-nvidia-blackwell在 X 查看被引用的帖子
来源：MiniMax (official) · x.com