Qwen3-8B v6e-8 — MFU & throughput per experiment
Pick MFU or TPS · all MFU on the causal basis (÷2 attention) · line = running-best frontier · ○ 8k / △ 2k · dashed = MaxText (MFU causal-adjusted: 8k 39.8% / 2k 36.6%; the tpu-recipes-v0.1.4 figures 45.3/38.0% are non-causal) · dotted = MaxText onset · click a point to open its experiment page · drag to zoom