At this point it's easier to just adjust everything to die sizes to get where each GPU "should" go (or how much NV is making money from them).
Ada (2022)
[Name] : [Die size] | [relative wafer usage]
AD102 : 609mm2 | 100%
AD104 : 294mm2 | ~25%
AD106 : 188mm2 | ~15%
(keep in mind this is AREA, so you can actually fit ~4x AD104 chips into the same space as 1x AD102, also I don't into account yields or PCB/cooler at all... also, also - making/designing/developing bigger chip
may be harder/costlier than smaller)
In another words : You can make 188x AD102 chips, OR around 609x AD106 (assuming perfect yields on both cases).
Here's "profit" comparison (I count only full die since NV pays for all silicon, regardless if it works or not) :
[AD102] 188x 4090 (MSRP 1600$) =
300 800$
[AD106] 609x 4060 Ti 16GB (MSRP 500$) =
304 500$
^That's how to make Money, BUT...
---------------------------------------------------------
[AD102] 1x 4090 (MSRP 1600$) =
1600$
[AD104] 4x 4070 Ti 12GB (MSRP 800$) =
3200$
NV, will profit at least 2x more from selling a 4070 Ti than a single 4090 from pure die perspective (again, yields for full core should be lower than "hacked" one from 4090).
Ampere (2020/2021) :
GA102 : 628mm2 | 100%
GA104 : 392mm2 | ~62,5%
GA106 : 276mm2 | ~44% (NV can make ~2.25x more GA106 vs GA102.
Pricing is... meaningless since you couldn't buy those cards at MSRP back in 2020/2021.
Regardless, from wafer area usage - you can make either 69x GA102, or 157x GA106 (again, perfect scaling on all wafers).
In which case :
[GA102] 69x 3090 Ti (MSRP 2000$) =
138 000$
[GA106] 157x 3060 12GB (MSRP 330$ =
51 829$
So, in this case selling RTX 3090 Ti was vastly more profitable than 3060s (again, assuming yields were perfect [they were not], and 3060 12GB uses slightly cut core spec so real yields get boosted a bit).
Way different than almost the same values on Ada.
Turing (2018/2019) :
TU102 : 754mm2 | 100%
TU104 : 545mm2 | ~72%
TU106 : 445mm2 | ~59%
"Wafer area adjusted" :
[TU102] 445x RTX Titan (MSRP 2500$) :
1 112 500$
[TU106] 754x RTX 2070 (MSRP 500$) :
377 000$
^Again, RTX Titan was WAY more profitable from NV standpoint than mid range card.
Pascal (2015/2017) :
GP102 : 471mm2 | 100%
GP104 : 314mm2 | ~67%
GP106 : 200mm2 | ~42,5%
"Wafer area adjusted" :
[GP102] 200x GTX Titan Xp (MSRP 1200$) :
240 000$
[GP104] 300x GTX 1080 (MSRP 600$) :
180 000$
[GP106] 471x GTX 1060 6GB (MSRP 300$) :
141 300$
In Pascal case I adjusted all cards (assuming wafer area price usage is the same), the more performance you need the more NV get's payed (and this
SHOULD be the norm)
And 900 series (2014/2015) as it's last 40nm (on
very mature process) :
GM200 : 601mm2 | 100%
GM204 : 398mm2 | ~66%
GM206 : 228mm2 | ~38%
"Wafer area adjusted" for all cards (ie. amount dies made is equal in wafer size consumed) :
[GM200] 45 372x Titan X(M) (1000$) :
45 372 000$
[GM204] 68 514x GTX 980 (550$) :
37 682 700$
[GM206] 119 599x GTX 960 (200$) :
23 919 800$
Again, we can see that most profitable was Titan...
This was a good math exercise
I hope I didn't made an error, but just wanted to check what this will come up with.
Maybe someone finds this interesting.
Sadly, I don't know die area number for Blackwell stuff yet :/
