Apple's biggest bottleneck for real-world inference is prefill processing.

Much less true since the M5 generation. Prefill, aka prompt processing, got a 4x increase.