Yarrow is correct in that AMD/Nvidia could/should be providing us with superior technical advances... Well, Nvidia should be anyways...
Why?
Nvidia, nor AMD, are responsible for the process node cadence of pure play foundry companies.
Both Nvidia and AMD are nearing the absolute limit of TSMC's manufacturable die size of 625mm², and no other foundry is presently capable of producing competitive GPU's, and nor has been for some years since UMC and IBM offered anything remotely competitive.
Both Nvidia and AMD's designs have transistor densities far higher than any other large die IC despite devoting a larger part of their IC's (around 50%) to low transistor density uncore ( memory controllers, cache, I/O). Typically a GPU's density is 12-14 million per mm². That density has only just been exceeded by Intel on 14nm - a full node smaller than TSMC's 28HP.
The flaw is, I suspect, that since AMD can't keep up, Nvidia has no pressing need to improve either... as long as they stay ahead of AMD, they can charge whatever they like...
No, the flaw is that margins are razor thin for most discrete graphics cards. Most production/sales accrue from entry level and lower mainstream market segments. Enthusiast/performance desktop and professional graphics are sold in relatively small numbers (albeit at high margins), but every large GPU is a sizeable investment. If it doesn't meet all of its goals (performance, competitive (halo) leadership, sales numbers and longevity) then the design can wreck a company. Take Fiji for example - AMD gambled on a doubled up Tonga being more competitive against a 50% larger GM204. With the title "Worlds fastest GPU", AMD would have sold a boatload of R5, R7, and R9 300 series cards - and you can bet that they wrung everything they could out of the GPU in order to achieve that goal. Yet the company still fell short, and the overall impression is "nice try", not an awestruck audience. Big gamble which should be shown by AMD's next couple of financial quarter reports. Hundreds of millions of dollars spent designing a GPU that has to exist (probably) solely as a Radeon card, since I doubt there are many FirePro options available for 4GB framebuffer, 225-250W power envelope, air-cooled/passive workstation/HPC offerings.
Last year, a grand total of 51.39 million discrete cards were sold worldwide. A huge proportion of sales coming from entry/low performance video out cards sold by OEMs and those available for less than $100-139 retail. This is six million less than 2013, twelve million less than 2012, and sixteen million less than 2011. The market isn't huge for serious (performance/enthusiast) graphics in global terms - if it were it wouldn't be fought between two companies that are relative minnows in the tech industry. Expecting quantum leaps in performance when the market is most definitely constrained, the foundry processes are limited (to say the least), and the graphics pipeline API's advancement is of glacial slowness is unrealistic in the extreme.
