Quote from: Duds on January 17, 2023, 04:01:51It is a common mistake to assume that AMD or even Intel can just make their integrated GPU's faster. They are limited by memory bandwidth not by the size of the GPU.
Switching to DDR5 gave APU's a really nice generational bump but there'll be a bit of a wait for a new memory standard.
You can add more cache or you can add HBM but both of these massively add to the cost and APU's are supposed to be the budget option.
There's surely more gas left in the tank after Rembrandt/Phoenix. And maybe that will have to be assisted by faster memory. Rembrandt officially supported DDR5-4800 and LPDDR5-6400, but we've already seen talk of LPDDR5X-8500, LPDDR5X-10000, and DDR5-12600, for example. DDR5's changes to power/voltage management are going to allow massive increases in speeds over the entire generation. LPDDR5 already has its successor ready for 2023-2024 products.
I really want to see big cache and L4 on APUs, but it's a departure from what AMD wants to do. If they can design monolithic dies to have TSV connection points, maybe they can put an optional cache on only a top model, such as the equivalent of a Ryzen 9 6980HS or Ryzen 7 5700G (which was not originally a budget option with a retail MSRP of $359). Only a fully enabled and highly clocked model would need all the memory bandwidth anyway, not the models with half the CUs.
I don't believe AMD should be expected to go in this direction, but if they want to replace low-end discrete graphics on mobile/desktop with APUs, they have ways to do it.