Modules B02–B03 · concept player
The roofline, with the dot in your hands
One formula draws the whole curve: attainable performance is the minimum of the machine's peak arithmetic rate and intensity times its bandwidth. Slide the intensity and watch which term wins — then move the roof itself.
Illustrative peak, TFLOP/s: 100
Illustrative bandwidth, TB/s: 2.0
Arithmetic intensity, FLOP/byte: 1.0
Common workloads
5.0ridge point, FLOP/byte
10.0TFLOP/s attainable
BW-boundwhat limits you here
10%of illustrative peak reached
Log–log roofline
Peak and bandwidth here are illustrative, not any vendor's spec. They exist to let the roof itself move so you can see the ridge point shift. Published ceilings are a starting point, not an answer — B02's lab derives real peak FLOP/s and bandwidth from your own hardware, and every kernel this curriculum writes plots left of the ridge on real silicon (B02 section 5), which is the argument for fusion in one picture.
