精选理由
MiniMax搞了个M3内核，KV-stationary让稀疏注意力在B200上跑到980 TFLOP/s，每个块只读一次，长上下文推理加速神器。
MiniMax AI发布M3内核，针对长上下文稀疏注意力中数据依赖块选择导致的内存访问瓶颈。M3采用KV-stationary设计，每个块仅读取一次。在B200 GPU上达到约980 TFLOP/s的算力。该方案有效提升长上下文推理效率。
原文 · Fireworks AI
Long-context sparse attention has a catch: data-dependent block selection wrecks memory access kills...
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 → fireworks.ai/blog/kernel-op… 💬 0 🔄 0 ❤️ 10 👀 3325 📊 2 ⚡