[nine_k]

Specifically the M series from Apple have a very wide, very fast interface to DRAM, which is connected to DRAM chips soldered basically next to the CPU. That makes it possible to use the entire unified RAM as the GPU RAM, and reasonably run decent ML models (for code, text, audio, pictures) locally. No CUDA, no kilowatt power supplies. This is the real differentiator.

[cromka]

Notably, this is also Qualcomm's X2 architecture.

[eqvinox]

> That makes it possible to use the entire unified RAM as the GPU RAM, and reasonably run decent ML models (for code, text, audio, pictures) locally. No CUDA, no kilowatt power supplies. This is the real differentiator.

That might be relevant and a differentiator in your circles; it is entirely irrelevant in mine. Plain basic integer performance wins here.