Roofline attainable-performance bound
Bounds attainable kernel performance by whichever ceiling is lower: compute throughput or data-movement throughput.
Attainable performance equals the minimum of peak compute performance and memory bandwidth times arithmetic intensity.
A kernel runs no faster than its compute hardware or its ability to feed data, whichever limit is reached first.
Minimum is the main verb: performance is capped by the lower of two independently calculated roofs.
The Roofline model is an upper-bound diagnostic, not a latency prediction; overheads and imperfect utilization can place real performance below the roof.