AMD just seems to be stuck - they've invested in so many pointless false starts and long term plans that they just cannot keep up.
Tonga - Backported to 28nm after 20nm crapped out. Regression in efficiency and performance compared to Tahiti while having a bigger die. At least this is true for the 285.
HBM - 2 years too early to make a difference, undoubtedly contributed to making Fiji prohibitively expensive.
DX12 - ACEs in 2012 card. Still waiting to be used mid 2016. Maybe this will eventually pay off, but it came at the cost of DX11 performance for over 4 years and counting.
Fiji - Worthless chip for AMD. Costly, terrible yields/supply, and its competitor was so obviously better it's a joke. They would have been better off scrapping it and just letting Hawaii and Tonga stick it out against Maxwell. From an design standpoint, it's actually quite acceptable/good and HBM is impressive, but not nearly enough.
Global Foundries - Their production of Polaris has been exemplary (miraculous) relatively speaking, but TSMC is so far ahead, at least in performance and efficiency, it's sad.
If any of this really paid off, especially Tonga and DX12, they may have been able to surpass NVIDIA, but NV had the foresight to realize that neither of those things mattered and stuck with proven tech and current needs. All this 'orthogonal' R&D has been worthless and look what they're trying to do with Polaris - backtrack. Color compression, tessellation performance and DX11 were all meant to be addressed by this chip, the one coming out in 2016, when they should have been addressing these things 4 and a half years ago - instead we got DX12 preparedness. They are way too forward thinking. By the time the 7970 beats the 970 consistently, both of those chips will be obsolete.
Of course, it's too little too late. Tessellation performance is still bad, color compression is below Maxwell, and DX11 efficiency is taking baby steps.
Edited by Pyrotagonist - 6/29/16 at 11:12pm