Here is a demo showing running 1080p at medium (a few settings low) without mantle. Without mantle. And mantle is what makes this really interesting.
http://www.youtube.com/watch?v=O07YOk4nLyo
I think its perfectly fine to asume this can be done at 720/768p at high on the same machine. With mantle this could be doable in a mobile form factor as the cpu is reliewed and memory bandwith is the limiting factor anyway. Remember kaveri tops out with less tdp on the desktop size compared to richland.
I think we will see this apu eradiate all the mobile dgpu 640/650 class hardware if not 660 class already 2014. Remember an competitive intel solution needs to be quad core to do the same on the new games using mantle. Kaveri will have a huge cost advantage for the oem.
Comming from richland to kaveri + mantle in a game like bf4 is a gamechanger as amd is then no longer cpu limited. Kaveri will clearly then be heavily gpu limited.
Ofcource the lack of ddr4 probably changed the kaveri arch to have fewer shaders and probably half gpu power. But the upside to this is its a better fit to notebooks on the current 28nm gh hpp process because of tdp limitations. And its cheaper too, than what could have been twice as fast on the gpu side. Actually i think amd is more lucky than unlucke here than they think. The market is for the low end. They need 20nm for 1024 shaders anyway.
When the dust settles (and price falls - preordering prices is as always bs) after jan feb we will see tons of cheap 15 inch 720/768 machines using this apu.
The desktop specs and pricing compared to richland is here:
http://wccftech.com/amd-kaveri-a10-7850k-a10-7700k-preorder-prices-revealed-listing/
But the market for kaveri is on the mobile side, we will have to wait for that
Edit: bf4 without mantle on kaveri with different resolutions and settings (2133 mem)
http://www.youtube.com/watch?v=axyHkKn_e80
Edit At 2:20 the are playing bf4 at 720 on high - without mantle
- damn they even get about 25fpus at ultra without msaa at 720. Lol. Thats your proof right there !
(I havnt seen this before and i am impressed - there must be some secret memory efficiency in the cgn 1.1 arch to achiewe this - this is crazy and then only using 2133 mem, this begs for faster memory and a solid overclock)