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 inc

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.