You can't mention local AI on x86 without mentioning Strix Halo. LLMs are all about fast RAM. VRAM is the fastest. RAM is relatively slow. On this laptop, with double channel RAM, it's about 90gb/s, VRAM is up to 1.7tb/s on a proper 5090. Strix Halo has 128GB of quad channel RAM, which is about 256gb/s.

If the model fits in VRAM, it's the fastest by far, if it doesn't, it gets a lot slower. With Strix Halo, larger models can still be run at tolerable speeds for many purposes, but with dual channel RAM at half that speed, things slow way down to the level of unbearableness. The breakthrough models today barely fit in 24GB VRAM, other models are niche or neat toys.

Someday, consumer x86 will start to match shared RAM speeds higher end Macs have had forever, until then, sites like notebookcheck will keep glossing over that elephant in the room.