Ok, we'll just agree to disagree here. You also didn't account for next gen games taking advantage of more demanding features (Asychronous compute, Voxel global illumination, etc.) that will run faster on newer architectures. Sure, if you want to derive your averages based on 1000s of games, then you may be correct since we have had many years of straight up console to PC ports using outdated game engines. Of course it's not realistic to expect the new GPU architecture to run older game engines 80-100% faster since they don't have the latest tech and forward looking game code. Secondly, if you focus on 1920x1200 or below benchmarks, sure newer GPus won't be as dramatically faster since they'll be CPU bottlenecked.
What about newer games? 780Ti
gets killed by
Maxwell.
Don't you think that games released in 2016-2018 will use more demanding graphical effects, perhaps next gen effects that run much faster on newer architectures? You are also assuming that NV will focus just as heavily on Maxwell driver optimizations in 2016 and beyond which is a bold assumption after what happened with Kepler. Driver optimization alone could give Pascal a 10-20% advantage over Maxwell in 2016 and beyond titles.