Each GPU is a roofline (memory-bound ramp into a compute-bound ceiling). Each workload drops a vertical line at its arithmetic intensity. Hover an intersection for the attainable-performance breakdown.