If people don't mind a short rant, I want to make some comparisons about Maxwell and GCN 1.1 and why I think both of them are good in their own ways, because I think we sometimes lose sight of the fact that both camps are giving us some great hardware.
The thing is that AMD and nVidia cards are two very different animals even though they're doing the same thing. On one hand you have Maxwell in the green camp that packs a whopping 128 stream processors per SMM (11 SSMs for the 970,) versus AMD in the red camp with GCN 1.1 like on the 390 and 390X which have 64 stream processors per CU and 40 CUs like on the 390. Without getting too discouraged however, we've seen that AMD cards seem to have some muscle when it comes to GPGPU and parallel compute, probably a result of the large number of CUs and the really wide memory bus but, doesn't this sound a little too familiar? This is similar to the problem AMD had with their CPUs compared to Intel. AMD decided to give up some single threaded performance to get relatively large gains on parallel throughput. For a GPU this makes more sense than a CPU but when push comes to shove, you can only do so much in parallel, even more so with libraries only starting to become more parallel in nature. This is where nVidia wins because the beefier SMM cores in Maxwell and higher clocks simply produces more performance in most gaming situations, there is no denying that in my opinion and nVidia should be commended on a job well done. The more serial (not parallel,) the workload becomes, the better nVidia does. On the same token, applications that execute more GPU instructions in parallel where the rendering can actually utilize more of those 40 GCN 1.1 CUs like in the 290/390 and you have a formidable product. Problem is that not all games are coded well in order to use parallel resources and as a result, nVidia, like Intel, wins because their cores are faster and can do more with a focus on what's needed most with current games however, that gap evaporates when you start utilizing the extra resources that GCN 1.1 has despite the lower clock frequency at the cost of dev hours and more complex code.
So this really comes down to how games are developed, what kind of calls are being made to draw and render scenes, etc. So it's not just a matter of ROPs, or TMUs, or plain numbers. It's a matter of AMD going above and beyond the parallel compute thing but losing sight of where the market really was at. Unlike their CPUs though, I think that all of that work to add more parallel throughput to their GPUs is going to pay off as games start demanding more GPU power and we start seeing APIs like Vulkan and DX12 take advantage of it which is starting to better utilize these untapped resources.
I hope my rant wasn't too off topic.
