Lafiel said:
Yea, large monolithic chips hopefully can be replaced by clusters of chiplets in the future as that would mean much less wasted silicon waver area and should result in better prices. The problem is the bandwith of Infinity Fabric (IF), which afaik isn't quite high enough to seemlessly link several GPU chiplets into a big GPU. Even if AMD has a much more advanced IF rdy to go it could be prohibitively costly, as they don't seem to use it in their upcoming Navi GPUs (the one Lisa showed on stage at Computex was monlithic). |
I was also thinking about AMD's GPU roadmap and how Navi was said to have next gen memory (GDDR6) and scalablility. While that could mean different things I wondered if it could potentially hint towards multiple GPU chiplets like the Ryzen CPU chiplets.
There are other rumors at the moment that Navi 10 is still partially GCN and partially RDNA, and that Navi 20 next year will apparently become fully RDNA. With AMD mentioning that PS came into the picture after Navi had been in development, maybe the reason for the hybrid GCN/RDNA was for them? Since PS can add future advancements to their planned past/present GPU architecture using AMD's semi custom branch, maybe it's possible that PS5 could use smaller multiple APU chiplets based on Navi 10 that incorporate a Navi 20 advanced IF chiplet design?
The advanced IF could cost more but likely shouldn't be more costly than fabricating an 8 core, 72 CU, monolithic APU, and it's losses due to yields. If it works for CPU and can have costs as low as they do, then surely it could work for GPU as well once the die get's large enough that yields are a problem. Could it work for smaller separate APU's though?