You'll never be able to be as efficient as a monolithic chip, going off die costs power. In the paper was a good table which shows this:
On Chip you only need 80 femtojoule to move a bit,
on package you already need 6x more. There might be ways to improve this a bit, but it will always be worse than on chip.
It's more an economical need. You don't have another possibility because the smaller nodes need too high R&D costs. AMD managed to make 2 Polaris last year without a high end gpu and this year probably they will only bring the higher end and not refresh the mainstream. They don't do it because they like it, but because of not enough money for R&D. On 7nm R&D cost should be about double than 16nm. That would make the situation even worse, but by building 1 die and scaling it to 2 or even 4 multidie gpus you save soo much money. With Navi i'm pretty sure AMD will have a top to bottom lineup if they really use multidie gpus. Additionally you have smaller chips which means better yield. Nvidia on the other hand with single chips will have a power advantage, but higher production and r&d cost. But as on 5nm and smaller the r&d cost are skyrocketing, they will have no other way as to do the same.
@zuzu
Multidie concepts aren't like sli. You have enough bandwidth between them, that they behave like 1 chip.
But all this MCM stuff is further future. Here we need more volta news.
Facebook
Twitter