tangosmango

It does better under demand titles, and esp better with higher resolution/more intense textures (even though personally I wouldn't use it for anything other than 1080p). And games only get more intensive. It does have power drawbacks though sadly.
 

I mean technically it's here now. But no current games will use it, and the first games coming out are this year on it (and on a relatively limited scale). It won't be every game (or even a majority of games) likely until late 2016 or early 2017, by then both Nvidia's new lineup and AMD's new lineups will be out.

AMD can explain it better than I can. Really it's just allowing the GPU cores to work on multiple things independently. They use the massively parallel compute engines to schedule tasks asynchronously, so that no part of the GPU is waiting on another. It just aggregates the workflow, something that the relatively serialized DX11 API hasn't been able to do.

 
Oxide's Ashes of the Singularity benchmark stacked a 290X against a 980Ti in DX11 and DX12. In DX11 the 290X was predictably far behind the 980Ti, but when put on DX12, the 290X shot up to within 2 fps of the 980Ti, and even pulling ahead at 4K. That's crazy. Of course, not the best example since it's only one game, but it shows the potential of DX12.

Even without Async (which GM200 also doesn't have, but it does well compared to the Fury X), something about GM204 and below just fails really hard at DX12. I don't know if its how much Maxwell compressed memory bandwidths and texture units or what it is, but GM200 does well enough compared to its competition yet GM204 and GM206 get #wrek'd .
Basically it allows higher efficiency for utilization by allowing things to be computed out of order. Maxwell is so bad at scheduling things back into out-of-order that thus far it's just been better to turn it off for Maxwell (but even the two games we have so far don't use it super heavily and the raw power of Hawaii and bus width really seem to be paying off.)

OpenCL performance, raw compute performance... I mean, the Async Compute is really where it's going to shine. Performance boosts of 50-70% when the new APIs roll around, perhaps even higher. It's going to wreck faces. That's what everyone is buzzing about.

Vram won't bottleneck at all at 1440p and even at 4k memory overclocking with compression allows for basically no notable bottleneck. That said a 390 is a better choice provided you can cool it. It is for now, and it massively will be in the future (NOT BECAUSE OF VRAM)

390 is slightly better for now, and down the road will rip the 970 to shreds, because the 970 can't do Async Compute like the 390 can.