The TSMC numbers are suspicious: TSMC has said N2 density is ">1.1X N3"; presumably that would be comparing N2 to N3B here; which is showing only a 1.03X difference.
It also looks like this leak is claiming Rapidus's "high performance" library (2HP) is comparable to TSMC's HD libraries (see below), which seems unlikely?
...
A lot of the numbers line up with the analysis here:
Reddit post says for HD Libraries:
18A = 185
Intel 3 = 140
TSMC N2P = 236
TSMC N3E = "216 or 192"
And then the post also lists HP libraries:
18A = 164
Intel 3 = 123
TSMC N2P = 197
TSMC N3E = 183
...
P.S. Some interesting text in the comments comparing Scotten Jone's projections for 18A/N3 density vs. these numbers:
For N3, I believe Scotten Jones uses the densest possible 1+1 config. I don't think that's appeared in any products yet, I don't think TSMC has shown it at any presentation, and I believe it's rumored to be coming with a possible "N3S" variant.
For 18A and N2, both of Scotten Jone's 18A and N2 estimates are ~30% higher than what I'm getting. This may suggest that there potentially is some scaling factor inherent to GAA that I'm not compensating for (hence why both GAA nodes are different by an almost equal factor), or that Jones is using a theoretical, or future "densest possible library" option on both nodes (18A-P potential denser libs, N2P finflex? or still, more dense libs) like he did with N3.
Funnily enough Scotten Jones's projection for 18A's competitiveness in density is worse than mine, he has N3 peak density being ~15% better than what my numbers show
