As for performance, an 1170 will match a 1080ti else no-one would bother buying a 1170 they'd just buy a 1080ti at it's reduced price. That's why a 1070 matched a 980ti and it's why a 970 matched a 780ti.
Surely most of the cautious comments here are based on a few major things:
- TSMC's 16nm to 12nm doesn't seem to offer any major density increases. A quick comparison of GV100 and GP100 shows around 3% increase in density
- We're assuming that the 'low hanging fruit' Nvidia found going from Kepler to Maxwell (taking out some compute functionality, moving to semi tile-based rendering etc.) for instance, are now gone and any architectural improvements will be much harder
Now, Nvidia surely know this so if they think they have to offer, for example, a 30% improvement it seems there is only one thing they can really do: sell a bigger die.
GP104 is only 312mm², while GP102 is only 471mm². This will affect Nvidia's margins but they may not have a choice. TSMC's 12nm seems to be okay for power usage if the Quadro GV100 is anything to go by
(https://www.anandtech.com/show/12579/big-volta-comes-to-quadro-nvidia-announces-quadro-gv100)
at almost the same frequency, GP100 only has a 15W higher TPU that buys around a third more die and around 42% extra SP and DP. And that's with 32GB of 16GB of HBM2.
So, assuming that Nvidia knew all this a few years ago, I think it all depends on whether Nvidia are willing to suffer a hit on margins.