Apple is in such an interesting position right now. Having made the decision for whatever reason to go with a unified memory architecture, they are now almost unique in terms of being able to offer super power efficient laptops with very large GPU memory. If they were to just double the memory bandwidth and do nothing else they could break away as a preferred way to run especially inference but maybe even training for small LLM models. Given the bottlenecks on availability and pricing of nVidia cards we are already toying with the idea of building a small local LLM stack based on Macs for this reason. If Apple made it more compelling we could easily go from toying with it to seriously considering it.

It's disappointing to hear they are delaying this year's Macbooks because I was really hoping to see some improvements like this specifically targeted at running local LLMs.