> My understanding is that the only reason dGPUs generally don’t use system RAM isn’t that they don’t have access, but that the supporting software doesn’t use it out of choice because the of the speed hit
It looks like one of the advantages of Apples put-everything-on-the-package strategy is that they can have a very wide bus to ram, which makes using system ram for the GPU much more palatable.
Isn't the main difference that Apple is building something like 4-16 channel memory controllers in a mobile device while you normally don't get more than 2-4 channels even on a desktop? That's a lot of transistors and (potentially) power usage but if you can get on the latest node and have a market willing to pay for giant chips it lets you get impressive amounts of bandwidth.
I don't think you need the RAM to be on the same package for that, it just makes the timings easier.
It looks like one of the advantages of Apples put-everything-on-the-package strategy is that they can have a very wide bus to ram, which makes using system ram for the GPU much more palatable.