Question Is it possible for NVIDIA/AMD to put two chips (die's) on one card?

Mar 19, 2019
10
2
15
0
Nowadays, aside from all the RTX cr@p, all Graphics Cards have 1 die which does all the fancy work.
This lead me to question and ponder ,whether it would theoretically be possible to just put a second chip (die) on the opposite side of the card / evenly across the card.

I mean I can't be the first person to ask this.
If the issue is:
  • Heat: Increase the height of the heat sink and/or insert other cooling solution
  • Sync. between both chips: CPU's already have technolgy implemented in them that allows for a more synchronous work so its possible for a Tech giant with $Bn to develop a solution to this ... right?
  • More heat: More Fans
  • Size: My Processor doesn't fill up the entire space of the case and we have standards , people! (ATX, ITX etc..) thus increasing the length or height should not pose any issues (excl. thickness/widening the cards size which could cause trouble with most cases, and we can't saw pieces of the glass panel out now can we?)
- More More Heat: More more cooling. Seriously, tech giants horde billions of dollars but pretend they can't figure those kinds of issues out...yea def...
View: https://imgur.com/ezTWwU5

Any reply appreciated, yes its a dumb question but its an idea so "it" won't care about any insults >> 💡✋
 
Considering that AMD is already utilizing multi-chip designs for their HEDT and server processors, and will be moving to a chiplet approach for their upcoming 3000-series CPUs, it seems likely that they might do something similar for future graphics cards as well. In the case of their CPUs, everything from their $100 budget processors to their multi-thousand dollar server processors will likely be using the same chiplets containing CPU cores, and only a secondary I/O chip will change between the high-end and consumer models. The 64-core Epyc processors will have eight chiplets, each with eight cores, while standard consumer parts will likely have up to two chiplets, for up to 16 cores.

In the case of a graphics card, they could likely do something similar, only with graphics cores. For example, they might have 16 CUs per chiplet. An entry-level card might contain one chiplet, cut down as necessary. More mid-range cards could utilize two or three chiplets, and a high-end card might utilize four or more. Of course, their main limitation would be power use and heat output, though they could likely achieve more efficiency by using more, lower-clocked cores, and a multi-chip design could allow them to be spread out a little more for better heat distribution. With a design like this, multi-card support wouldn't be a concern, since the chiplets would act together as a single graphics processor, all interfacing with a single I/O chip.
 
Reactions: deNameMo

ASK THE COMMUNITY