Associate
there have been reports that the tweaked 4NP or whatever they are calling it can achieve 30% higher density than the 4090 process, so right off the bat theres going to be approx. 30% more transistors even if the die size remains unchanged, 800mm2 is a bit too high..i will take any positive surprises that come my way, but i wont be personally betting on a 800mm2 chip.. and the rumour about number of shaders/cuda cores etc kinda help corroborate this assumptionYet TPU have their placeholder at x1.33 shader count and same 4N as ADA:
AD102 is 600mm², so x1.33 would be around 800mm², and seems is what is expect them to do while still on 4N. Maybe if they are separate libraries they could use, they can increase density a bit but they are hard to mix (the Zen 3D cache dies are produced separately using a tweaked process - unsure whether anything like this could be done on the same wafer as logic).
All the rumours have been pretty consistent on 4N though. I guess theoretically, Nvidia could go half chiplet like AMD by moving cache and IO to chiplets but no evidence of them considering that outside data centre.
Far more realistic is shaders 1.33 of previous, clocks back a bit since 800mm² chips are huge, but less disabled cores and faster memory for around 30% more performance.
Price? Well this would be the first time consumers get 800mm² which they usually reserve for data centre = £Not.cheap!
i feel its going to be more like 650mm2, because the process has matured since 4090 and its cheaper than before as the tech has moved on to a new cutting edge 3nm node, so i believe theres going to be a slight bump to the real estate
Last edited:
