The irony now is that NVIDIA need to throw away a lot of that efficiency out the window in order to hold off Intel from making headway into the HPC market.
This is what Pascal will achieve by re-integrating those FP64 units NV threw out with Maxwell. Come Volta, we'll see even more hardware redundancy added to NVIDIAs architecture (potentially re-introducing hardware side scheduling if that's already not going to make a come back with Pascal).
I also expect larger caches in both Pascal and Volta. Further eroding that efficiency lead people praised Kepler and then Maxwell for bringing to the table. We may also see dedicated Asynchronous Compute engines making their way into Volta.
Basically, NVIDIA are heading towards a more GCN-like architecture due to pressures on both the supercomputer end as well as the consumer end with DX12 and Vulkan (let's not forget VR).
AMD will be refining GCN with Polaris and Vega. Introducing efficiency with architectural tweaks aimed at boosting IPC coupled with a 14LPP process.
AMD are chasing VR and looking to also make a return to the HPC market with Zen, high bandwidth point to point interconnect, CUDA to OpenCL conversion and Polaris/Vega based Firepro's.
It will be interesting to see how the three companies look by Q1 2017.
Im sorry if I missed the point or your intentions but I cant help thinking just how this post sounds like a pure AMD PR spin.
The addition of FP64, NVLink and other GPGPU/HPC related features will most likely feature only in GP100 and whether or not they come enabled in their consumer spin-offs is yet to be seen. Plus double precision floating point does not benefit 3D gaming at all while increasing overall power consumption.
Now last time I heard, the GK210 isn't that bad compared to the latest intel offerings. Furthermore the next generation gets to see a massive node jump where as previously IHVs have been stuck at 28nm so they gone out of there way to do all they can to improve performance while relatively maintaining similiar power consumption to their previous generation(s) of products.
When I look at nVIDIAs GPU advances from Fermi to Kepler to Maxwell, they've made some really good progress technologically. Pascal will add to this seeing as it won't be a completely you uarch. They have their own recipe for success (and this shows in their GPU market share and financial performances) and I dont see them venturing off to something different.
As a consumer, I do get frustrated that old generation products aren't optimised for new games, or perhaps the GPU itself wasn't forward thinking enough but looking at it from their business and engineering perspective its good because it leaves open for more improvement with their next generation of products. By the time we need the
actual performance for native DX12 games, new products that are much better at similar price are their as an upgrade path. This performance talk of DX12 this early in the game sort of reminds me of the SM3.0 vs 2.0 debate for example.
P.S. I really do hope zen will be successful. Im sure Polaris and Vega will provide serious competition to nVIDIA but for AMD, zen needs to atleast go close or be par with Intels last gen offerings. But this is OT.