Fermi design was set in stone before all those troubles (yield problems, leaking, thermal performance, TSMC, etc.) arised. Nvidia had a good idea how it would turn out but it's impossible to be 100% sure until you have actual silicon in your hands. But they couldn't go back to the drawing boards once they realised they're in trouble, this takes years. There's a thing called time to market.
Fermi was supposed to have 512 working cuda cores, it's not like Nvidia said, "Uhm, OK, 480 cores is just fine". All those problems combined resulted in this "catastrophic" failure (like an air plane crash) and Nvidia had to do a lot of adjustments to release the gf100 as we know it now.
Bookmarks