Soldato
Kope's claims for GB202, which would be used for a full fat rtx5090/ti
* Same TSMC 4NP as data centre cards. Details about this node is not public, don't know how much density or performance gain over last time's standard 4nm
* 12 GPC units per SM, same as Lovelace
* 8 TPC per SM vs 6 for Lovelace
* Significantly Increased L1 cache for SM throughput
* Total SM 192: 24,567 Cuda cores vs 16,384 cores on RTX4090
^^ Specs are for a full size GB202 die, which may not be the full size used in an RTX5090 due to cost, size and power.
NVIDIA RTX 50 "GB202" Gaming GPU reportedly features the same TSMC 4NP process as B100 - VideoCardz.com
Gaming GB202 “Blackwell” GPU using the same node as data-center B100 Kopite confirms B100 and GB202 share the same process. NVIDIA unveiled its latest Blackwell GPU architecture designed for data centers and AI acceleration yesterday, marking the first step in their new architecture push...
videocardz.com
I was surprised GB200 are based on TSMC 4NP instead of N3 so Blackwell gaming chips will use same process as GB200. Kopite claimed N4P has 30% density increase compared to old N4 process.
I think GB202 will be an interesting chip with 192 SM, 24.567 CUDA cores and 512 bit bus, Kopite not sure GB202 will be MCM chip so looked at GB200 it used dual GB100 chips so maybe GB202 on RTX 5090 will be MCM chip with dual GB203 chips. RTX 5080 will use single GB203 chip with 96 SM, 12,283 CUDA cores and 256 bit bus, maybe Nvidia will reserved 384 bit 1 year later for Blackwell refresh for cut down RTX 5080 Ti or SUPER.
Last edited: