The problem appears to be with TSMC, rather than AMD or Nvidia. I'm sure both would have had 20nm done months ago if TSMC was on schedule with their original 20nm predictions.
Certainly the slow down with transitioning to a new node is largely at the hands of TSMC, but the point I was making is that AMD has been designing two different sets of chips for the past few nodes and this type of workload may be catching up with them. At 40nm, they made Cypress/Evergreen/etc. then made Cayman/Barts which were not insignificant redesigns. Now at 28nm, they made Tahiti/Pitcairn/etc. and are preparing to release new chips which, again, are not insignificant in their modifications. They are not taking existing chips and tweaking a few transistors like Nvidia did with Fermi going from the 400 to 500 series, they are creating altogether physically different chips which takes way more R&D than simply tweaking existing GPU's.
Nvidia, on the other hand, released one set of chips with 28nm and while they did release significant refreshes with 40nm Fermi chips, all of Fermi v2 chips were tweaked but otherwise based entirely on their counterparts that were replaced and did not differ much (if any) in transistor count or die size. AMD may be executing smaller, incremental design changes between each new family of chips when compared with Nvidia's, but I'm arguing that it ultimately takes more resources to release 2 sets of different / overhauled architectures totaling 8 chips on the same node vs. 5 off of one architecture.
(I researched and added it up here:
ANALYSIS. I found that AMD released 8 physical GPU's at 40nm, while Nvidia released only 5. So far at 28nm, AMD has released 5 so far with obviously more coming.)