That FPGA thing ...
I have no figures, but my gut feeling is that there are quite a few applications where a relatively small amount of custom hardware would make a phenomenal difference, particularly if it was really close to the CPU. (The "next" video compression standard, where "next" is "the one after this chip was made" springs to mind, but cryptography and even funny kinds of string searching are probably fruitful areas to explore.)
So in 5 years time are we all going to look back at 2016 and think "Gosh, they were still wasting 2/3 of the die area on a GPU that was old-school before they'd even finished the design. What fools they were!" ?