5 · Serve and the GPU lie · 31/38
Roofline
Two ceilings. Intensity picks which.
playGPUDrag the probe, or click an op. Left of the ridge = memory-bound = a fraction of peak.
H100 ridge 295 FLOPs/byte. Prefill sits right. Decode of one token sits at I≈1 — HBM. Same weights.