I don't share your optimism. More bandwidth doesn't mean more TFLOPS. By the time 14nm is ready next year, 28nm is more than four (yes FOUR) years old. If history repeats itself, we'll still be on 14nm in 2020.
Well, a little bit more optimism would be warranted if you had done some research. The reason we'll have been on 28nm for four years is because 20nm was a bust, and was only used for SoC (aka iPhone 6). But it's not a situation that's likely to repeat again. 20nm was the last planar node for TSMC, and was basically created to ease the transition to FinFET transistors. For this compromise, it had all kinds of issues like poor power scaling.
However, 28nm --> 16FF+ will be like a node shrink+, with a 2x density improvement (instead of the usual 1.9x) and 70% (!) less power consumption - that's normally like 40-50% at best.
In their Q1/15 report, TSMC stated that their 10nm qualification is slated for late 2015, with volume ramp beginning at the end of 2016. Granted, there's still room for slippage and the first 100 million chips will be SoCs, but expect 10nm GPUs roughly a year from then. And it will be a true node shrink, too, with 2.1x the logic density and 40% power reduction.
As for 7nm, risk production is targeted at early 2017. The longest cadence in TMSC's history from an announced risk production date to product in the shelves is about two years--and that's the worst case scenario with lots of delays. All this is to say, there will be 7nm GPUs in late 2019/early 2020 at the latest.
GF110 (2010, 1.5 TFLOPS) -> GM200 (2015, 6 TFLOPS), a 4x improvement in 5 years.
That's pretty impressive with only one node shrink, tbh. It would only yield a 2x improvement in this case, which means the underlying microarchitecture has also been progressing. Also, Maxwell flops are more valuable than Fermi flops.
Assuming PS5/XB2 will use a budget GPU again, we'll get 7 TFLOPS consoles in 2020. (1.8TF x 4). Basically the power of a Titan X, but probably with a lot more memory and bandwidth (HBM).
In fall 2013, the most powerful GPU on the market was the R9 290X, which was 5.632 / 1.843
≈ 3x more powerful than the PS4. AMD will probably release a ~50 TFLOPS GPU in the 2019-2020 timeframe (The 390x this year will be around 8 TFLOPS, then we'll have three die shrinks, so 2^3 * 8TFLOPS = 64 TFLOPS). Taking the conservative 50 TFLOPS number and dividing it by three, a PS5 released circa 2019-2020 should easily be around 15 TFLOPS--which is as it should be, cause it doesn't make any sense releasing a console that's not close to an order of magnitude more powerful than its predecessor.