From my perspective, working just on compute, it looks very different:
7950 five times faster than GTX670 (although the latter was a bit faster in games)
280X two times faster than Kepler Titan.
GTX1070 exactly in the middle between 7950 and FuryX. (interesting: Here real world performance of 1 NV TF matches 1 GCN TF for me)
(no data for generations in between and more recent stuff)
But NV always won all the game benchmarks, so i expect AMD to move from strong compute towards strong rasterization (and now RT?) as well. Major reason i'm personally always nervous with new architectures and less excited about HW RT than others
I see the TF value mainly as a measurement of general purpose compute power. It works much better for this than for rasterization rlated tasks.