|
|
|
|
|
by matthewfcarlson
1 hour ago
|
|
It's a pretty huge cost to support an entirely different set of hardware with different kernel extensions and an entirely different build (x86 instead of arm64e). Could apple choose to do that? Absolutely. But the cost of supporting an M1 is very different than the cost of supporting Intel. |
|
I suspect you cannot simply sprinkle AI functionality through an OS and manage the difference between unified and non-unified VRAM without noticeable tradeoffs.
The marginal impact of adding some tiny amount of foundational model use to an existing app function is very different between the two.
More so if you want to augment some existing functionality with model use, more so still if you were going to replace some functionality with model use (which I suspect is not yet happening).
You could do it if you were not concerned about surfacing the RAM/VRAM implications to the user through seemingly arbitrary clashes (worse graphics performance or not being able to use the GPU to process some video because you have the larger foundation model loaded, or an AI function refusing to run because another task has booked a lot of VRAM).
But Apple tend to be concerned about surfacing that sort of internal concept. Going forward with Apple Silicon alone means a bunch of questions like that simply don't come up.