Roofline model
Predicting performance from whichever resource saturates first — here, memory bandwidth.
A roofline estimate divides work by the limiting resource. For token generation the limit is reading weights from memory, so speed is bytes-per-token over achievable bandwidth plus a fixed overhead. It is accurate to a few percent for dense models on well-characterized backends and considerably less accurate for mixture-of-experts models, where far less public measurement exists.
Every speed figure on this site is a roofline estimate and carries its error band.