Fireworks AI 在 B200 上实现稀疏注意力 980 TFLOP/s

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. 👇