Bofferbrauer2 said:
PS4 came out with the 28nm process. It's followed by 22/20nm, 16/14nm, 12/10nm, 7/6nm and then 5/4nm. The OG Switch already had one and a half, while the HAC-001(-01) refresh had 2 node shrinks under it's belt compared to the PS4. The successor of the Switch will have at least 4 full node shrinks compared to the PS4. However, the closer we get to atomic sizes, the smaller the gains get of he die shrinks (especially for things like IO and cache), so it's not an exact science or very linear. |
By full node I meant quadratic scaling of 70% on feature size, not whatever name these companies' are giving to their product. Logic still shrinks perfectly and has nothing to do with diminishing gains (since smaller logic cells compensate for less scaling in SRAM and I/O) TSMC and Samsung are just not following the old ITRS guidelines.
20 nm is just one node below 28 nm and shares the same pitch as 16/14/12 nm, the only difference is that the latter uses FinFETs. 11/10/8 nm was the next full node below that. So... two nodes.
That's easier to realize by looking at transistor density, which was ~ 1.2 MT/mm2 for 90 nm, ~ 25 MT/mm2 for 20* nm (= 4 and a half doublings), and just ~ 50 MT/mm2 for 8 nm.
Last edited by haxxiy - on 14 November 2023