No we didn't, the last jump in performance with a new node was GTX580 to HD7970 and that was ~25-30%. Or 40% if we go from 6970 - 7970. GTX580 - GTX680 was a similar jump in performance. That's the usual performance jump when we go from full fat on old node to mid fat on a new node. Next gen Pascal and Polaris will start at mid sizes as is the new norm in this industry. These mid sized GPUs on a new node will bring the usual ~30%-40% performance increase and marginal power savings. It will be 2017 before we start to see full fat Pascal or Polaris. The only way I see a significant performance increase of ~50%+ would be if Nvidia release a large mm2 HPC pascal chip and salvage/harvest parts for a cut down consumer version. If that happens expect it to release at ~£900
Everything about this is wrong. 30-40% isn't the 'usual performance increase' at all.
5870 was 2.15billion transistors and 334mm^2, 6970 was 2.64billion transistors and 389mm^2 and 7970 was 4.3 billion transistors and 352mm^2.
So first off 5870 to 7970 was full fat to full fat. Second, the increase was around 60% for around a 1.9x more dense transistor part. IE 7970 literally doubled the transistor count and because of not being quite 2x transistor density it grew very slightly. It was around 60% faster but to a very new architecture which took time to bed in and the gains became much higher.
http://www.techspot.com/article/942-five-generations-amd-radeon-graphics-compared/page4.html
look here, today the 7970 offers over twice the performance of a 6970 and more than that compared to a 5870, much more.
http://ht4u.net/reviews/2013/xfx_r7970_double_dissipation_ghz-edition_im_test/index18.php
this review from 2013 shows the 7970 to be on average 92% faster than a 2GB 7970. So much of that performance gain was over the first year of it being available rather than the last couple of years.
On a given process node the density and performance(including clock speeds) you can achieve improve over time, what you can achieve even a year let alone 2-3 years later(think Kepler to Maxwell or compare Fiji to 7970 power efficiency) on the same process is very different.
The normal comparison is 4870 to 5870, to 7970, and you'd compare that with a circa 8.5billion maybe 320-340mm^2 14nm chip.
7970 at double the transistor count offered 60+% performance right out of the box for games that weren't designed or optimised for it, over the next year that increased significantly.
Full fat to half the size is not 'normal' at all. It's happened precisely once so far with a 520mm^2 gtx 580 to a 300mm^2 gtx 680. I mean you could say AMD did it with the 3870.... but not really. The 2900xt was never meant to be 80nm to begin with, the it was always supposed to be 65nm and had to be pushed back to 80nm which made it way bigger than they wanted.
For a given chip size, a new node offering near enough 2x transistor density has every chance of offering 60-70% more performance out of the gate without optimisation in games or drivers and 85-95% more performance within a year. That is what 'normal' is.
What is also normal is that huge chips are incredibly different to make on new nodes, Nvidia struggled with big chips at 65/55nm(only a little but the first nodes the yields weren't great at that size), at 40nm they had a disaster and at 28nm they gave up on making a big chip out of the gate which is why we had the medium chip first. 14/16nm should again see big cores out of reach and a likely longer delay between medium and large chips.