I've read that Maxwell/Pascal was a pretty nice architecture improvement on nVidia's side (with Pascal being more or less a shrink of Maxwell), and AMD has mostly been coasting on the same architecture for the past few years. This makes GCN pretty wild as an architecture (in theory you could probably have close to modern Radeon speeds with cards from three or four years ago) and a tragedy for AMD that their drivers lagged behind hardware that should have slam-dunked nVidia.
Maxwell and Pascal were definitely improvements over Kepler (which is comparable to AMDs x1000 in design and shittyness). But until they implement true async shaders and substantially improve their shader arithmetic I still think GCN is the overall better architecture.
Don't get me wrong, NVidia has made, and is making great cards. If I'd be willing to spend the money, I'd go team green. They are really good at the usecase of 'play current generation games'. They have the resources for a huge driver team which is essentially reverse engineering (often with help/partnership) games and try to optimise performance. That is a pretty big and expensive undertaking but NVidia is doing an exceptional job at it.
The big and expensive driver teams is part of the motivation for next gen APIs, to cut the team and raise profits. But AMD really has another problem that is deeper then driver level: The established libraries do not properly support their architecture. That is the other motivation for Vulkan (and it's successor Mantle); get a better intermediate language that can use more features of the cards.
Vulkan is also another try at dislodging MS by attacking dx, hence why Valve is such a big partner.
AMD has made some good incremental improvements to GCN since it's invention. But most of them are hardly used by many applications. But you are right, most of the extra power from modern AMD cards come from bigger chips and smaller processes.
We have a pretty good idea of what Intel's future chips will look like, and a decent idea of how nVidia will continue to design, but I don't know that anyone has a good idea of price/power on Zen or Vega, which makes it
difficult to make value extrapolations until we see them.
You are right, I hoped I made clear I was mostly speculating. Most info about Zen is hype and I highly doubt it will be able to truly compete with newer Intel chips for high end single core performance. It will probably be a little of a let down for some people. It still looks like it could shake the mid-end marked though, CPU prices are really high at the moment.
The newest rumors on videocardz is that Vega 10 is roughly around gtx 1080 performance in Ashes. That means behind gtx 1080 in most other games and ahead in stuff like madvr. Pretty much like the last high end generations of AMD. They still tend to price their cards rather competitively though and Vega 10 (according to rumors) might arrive before Zen does anyway. But I do not expect a major price shift in the GPU marked, the rx480 is already priced rather competitively and the 300$+ cards are largely irrelevant for most of the marked (except for the stupid 'NV has the fastest cards' hype).
We have a pretty good idea of what Intel's future chips will look like, and a decent idea of how nVidia will continue to design, but I don't know that anyone has a good idea of price/power on Zen or Vega, which makes it difficult to make value extrapolations until we see them.
Edited by Spock, 09 December 2016 - 01:13 PM.