I don't think LPDDR5 is that much more expensive than DDR5 on a per-GB basis. The performance advantage is mostly from a wider memory bus, and a bit from being soldered.
In other words, the CPU having access to huge amounts of memory bandwidth is not really costing the users much, after they've paid for the GPU and its memory system. The cost is really just the incremental cost of enlarging a GPU die by strapping some CPU cores on the side.
It may not be precisely most cost and performance optimized configuration for gaming, but on the other hand it provides huge benefits for anything that doesn't fit in a cheap GPU's VRAM (eg. LLMs).
> I don't think LPDDR5 is that much more expensive than DDR5 on a per-GB basis.
It is by the time the OEMs are through with it, and once you're soldering it you have to buy it from them.
> The performance advantage is mostly from a wider memory bus, and a bit from being soldered.
Even the little bit from being soldered should go away when CAMMs enter the scene. But the wider memory bus is part of what makes it expensive. It's not just the input pins on the APU, it's routing all of those pins through the socket and the system board and the memory DIMMs. Which you can avoid by putting some amount of HBM on the APU package without losing the ability to add memory via the narrower bus in the traditional way.
> It may not be precisely most cost and performance optimized configuration for gaming, but on the other hand it provides huge benefits for anything that doesn't fit in a cheap GPU's VRAM (eg. LLMs).
LLMs are nearly the only thing that currently uses this, and there is nothing stopping a CPU option with a large amount of HBM on the package from existing and fitting into the socket on an ordinary system board for the people who want that.
But CPU makers also don't really want to make a thousand SKUs with every possible combination of cores and memory, so what they're probably going to do is have lower core count CPUs with smaller amounts of HBM and higher core count CPUs with more. In which case you can still buy the $100 APU with a modest amount of HBM and then add as much DDR as you want via the slots on the system board instead of having to buy $1000 worth of CPU cores you don't need, or replace a perfectly good CPU with a different one, just to increase the amount of memory for the majority of applications that aren't bounded by memory bandwidth.
> I don't think LPDDR5 is that much more expensive than DDR5 on a per-GB basis.
It's also worth pointing out that while Apple is using LPDDR5 and getting more bandwidth out of it by just using more pins, that isn't that most GPUs use -- they use some version of GDDR, which is more expensive on a per-GB basis than DDR, because it has more bandwidth per pin.
So putting some "fast" memory on the APU package has the further advantage that you don't have to do what Apple did, you could actually put HBM or GDDR on the CPU package, and then the faster memory would be faster.
In other words, the CPU having access to huge amounts of memory bandwidth is not really costing the users much, after they've paid for the GPU and its memory system. The cost is really just the incremental cost of enlarging a GPU die by strapping some CPU cores on the side.
It may not be precisely most cost and performance optimized configuration for gaming, but on the other hand it provides huge benefits for anything that doesn't fit in a cheap GPU's VRAM (eg. LLMs).