=

The passage around this formula

The roofline model’s ridge point, $I^{*}$ = $P_{\max}$ / $\beta$ , says whether a kernel’s arithmetic intensity is high enough that memory bandwidth is not the binding constraint [ 7 ] . It says nothing about whether the array can actually deliver $P_{\max}$ once that bar is cleared, because $P_{\max}$ itself assumes every cell is doing useful work on every cycle — and a real systolic pipeline spends some of its cycles filling and draining rather than computing.

Read this part in the article →