I guess what matters most is performance-per-mm^2. If HBM3 is more than 3x faster then perhaps it's fine. Otherwise, it might make more sense for chip manufacturers to just add more memory channels or widen the bus width.

schmidtbag:

I guess what matters most is performance-per-mm^2. If HBM3 is more than 3x faster then perhaps it's fine. Otherwise, it might make more sense for chip manufacturers to just add more memory channels or widen the bus width.

Would it though? I'd assume most of these chips will end up in server arrays where density is not an insignificant performance metric

Yogi:

Would it though? I'd assume most of these chips will end up in server arrays where density is not an insignificant performance metric

Depends on the application I guess. Pretty much where HBM would be needed is if there's no room to add more chips to the PCB, the PCB is as large as it can get, both sides of the PCB are in use, and you still need more bandwidth and/or capacity. I'm sure such examples could exist.