Llama3-8B v6e-8 — MFU & throughput per experiment

Pick MFU or TPS · lane MFU is causal (÷2 attention) · line = running-best frontier · ○ 8k / △ 2k / ◇ 4k / □ 1k · dashed = MaxText 8k (MFU causal-adjusted to 39.6%; the tpu-recipes-v0.1.4 figure 44.6% is non-causal) · dotted = MaxText onset · click a point to open its experiment page · drag to zoom