Roofline Model
The fundamental formula that proves why decode is bandwidth-bound
Fundamentals
In plain English
A speed-limit chart that tells you whether your car is limited by its engine horsepower (compute) or the narrowness of the road (memory bandwidth).
Technically
An operational performance model relating attainable throughput P = min(Peak FLOP/s, Arithmetic Intensity × Memory Bandwidth). Demonstrates that prefill (high arithmetic intensity) is compute-bound, while single-batch decode (arithmetic intensity ≤ 2 FLOPs/byte) is heavily memory-bandwidth bound.