SemiAnalysis AgentX · B300 · TP4 on four GPUs

The specialized decode path recipe raises B300 interactivity by 44–59%

P90 interactivity (tokens/s per user) at concurrency 1, 8 and 16. Higher is better.

Before: Oct 2 recipe. After: Oct 20 recipe from InferenceX #4332.

Before #4332 (Oct 2) After #4332, specialized decode path (Oct 20)