The future

Epyon

[H]ard|Gawd
Joined
Oct 25, 2001
Messages
1,185
Do you think the gpu will go the same route as the cpu did.? They started with 2x dies then over 3 years went up to 8 dies. Do you think we with see the same growth over 4 years when they do land? I do 3d render work so I would hope they do not milk it with 2x dies then 2 years later 4x dies then 2 years later 6x dies. Seems kind of slow if you ask me. How do you guys think it will go down? I am kind of hoping low end is 2x dies and high end is 4x dies for like a titan class card. Just wondering.
 
Yes
  • Once the GPU can be broken up to chiplets, then I would think many different combinations will become available simplifying top to bottom designs with common components.
  • Appears, at least to me, Infinity Fabric -> way to connect separate dies effectively and Infinity Cache -> A way to connect different parts of a chip to cache data and instructions are related and will be expanded upon
  • Once the common items for any GPU, such as memory controller, multimedia engine, video out, L3 cache and using Infinity Cache maybe even exotic staging of using HBM, DDR 6 and maybe even Flash memory (have no clue what AMD will actually do) can be done.

Uniquely AMD diagrams for Infinity Cache would cover either CU's in a die or separate dies. https://www.freepatentsonline.com/20200293445.pdf
  • 210 is a interconnect (Infinity Fabric), GPU 204 single die with infinity cache, GPU 214 two dies connected or CUs connected and basically add as many CUs as you can connecting them together.
  • GPU 204 LO is shared L1 cache between the CU's, in GPU 204 example four are shared, in GPU 214 LO only shares two CU L1 caches
    • 2die.png
Looks like to me AMD has effectively mapped out a multiple die strategy for MCM designs for GPUs, how that plays out is anyone's guess.
 
Not so sure Amd will be doing chiplet before nvidia. I could be wrong. I am hopeful that 2x chiplet are entry-level and we get 4x on a upper end on its first release.
 
Do you think the gpu will go the same route as the cpu did.? They started with 2x dies then over 3 years went up to 8 dies. Do you think we with see the same growth over 4 years when they do land? I do 3d render work so I would hope they do not milk it with 2x dies then 2 years later 4x dies then 2 years later 6x dies. Seems kind of slow if you ask me. How do you guys think it will go down? I am kind of hoping low end is 2x dies and high end is 4x dies for like a titan class card. Just wondering.
This is the approach Intel is using with their GPUs.

What's not clear for me is if the OS will be able to see al the chiplets as a single GPU. Not an issue for AI, inferencing or data centers, but huge deal for gaming.
 
Back
Top