News GB202 die shot beautifully showcases Blackwell in all its glory — GB202 is 24% larger than AD102

Somehow, it feels strange to see the cache in big blocks, like that. I guess I'm so accustomed to seeing LLC being more distributed, but it's certainly better if a SM doesn't have to fetch something from the opposite corner of the die, for instance.

Here's the twitter link:

GiJNaMabUAA9p-O


View: https://x.com/Kurnalsalts/status/1883153126011892140/photo/1


I'd love to know something about its interconnect topology. I wonder if there's any direct GPC-to-GPC communication path, or if all GPC communication has to go through L2.

Also, does anyone know what "3D FF" is referring to?
 
Last edited:
Somehow, it feels strange to see the cache in big blocks, like that. I guess I'm so accustomed to seeing LLC being more distributed, but it's certainly better if a SM doesn't have to fetch something from the opposite corner of the die, for instance.

Here's the twitter link:
GiJNaMabUAA9p-O

I'd love to know something about its interconnect topology. I wonder if there's any direct GPC-to-GPC communication path, or if all GPC communication has to go through L2.

Also, does anyone know what "3D FF" is referring to?
I found a spanish forum saying 3D FF are texture engines/ RT Cores. For what it's worth. They were looking at AD102 vs GH100

 
  • Like
Reactions: bit_user