I think the 8GB of ram would be more limiting than the actual M3, or M2 even in the case of the new iPad Airs. The reason being that generative AI and LLMs and such are much more dependent on neural engines and GPU cores, not CPU cores. If you compare an A17 with an M2, technically speaking and in theory, the M2 is vastly superior on those two departments.
In fact, if u look how the latest AAA iOS/iPadOS games ( namely, RE and Death Stranding) run ion a A17 vs on a M1 iPad, the M1 is still superior. What does a gaming comparison have to do with in-device AI you ask? Well because AI, just like gamibg, depends a lot on GPU cores, so if A18 will be able run AI tasks that the A17 can't , it is reasonable to assume that the M1 might be able to run those too at a similar performance, meaning M2 and M3 will be better at it than the A18.
Which leaves the question of RAM . For you to run smoothly "your own GPT" in say, a modern Mac, an M2 and M3 are pretty ok actually, but the more ram you get the better. We are talking of about 24 GB)m minimum , (better 32) . I doubt they'll put more than 8GB on the base versions of the 2024 iPad and iPhones. Which leads me to believe that most AI features on them are not gonna be ran by the SoC in itself; it would be instead get to offloaded to Apple Servers.
We'll see.
