I found it funny to see the mini 970 thrown around as an argument. The fact that the 970 relative to the rest of nV's lineup isn't even top of the line makes it funnier.
AMD could take Tonga and cram it into a mini 970 PCB for all they care, considering the similar power requirements and both using GDDR5 connected with a 256 bit bus. That wouldn't be as impressive as what GM204 can do, anyway, under those constraints. Still, it'd be the perfect comparison.
On the other hand you have Fiji crammed into that ridiculously small PCB (nano) putting out performance comparable to the 980Ti with a joke of a heatsink, if you're inclined to set the power limit a little higher than 0% so powertune doesn't throttle it as often.
If nV/a third party manages to cram a full GM200 implementation in the PCB size of a Fury/Nano with an accompanying cooling solution, then we could have a decent comparison point. That doesn't seem feasible for them until HBM2 Pascal arrives.
Mini 970 can't be compared to Fiji (something enabled by HBM, after all) and its many versions with a straight face. 4 HBM stacks occupy almost the same space as one or two GDDR5 modules. Performance isn't even in the same category (even with Nano being such a "weird" product and priced as it is). It's apples and oranges.
-------------
HBM frees TDP that is now available to the GPU itself for improved performance apart from all the benefits we've already gone over and over and over in lots of threads around here relative to GDDR5 (and now GDDR5X). There's the interposer cost and increased complexity to pay for at first but it clearly is the future and makes a lot of sense for GPUs or whatever products that are enabled by it.
Just look at Fiji. It's self contained. You only have to add the power components and you're done. It's
elegant. No need to bother with all those GDDR5 modules and bus routing on the PCB anymore. Both Fiji PCBs are just that, connectors, sockets and power components in a compact package. An AIO/triple fan open air cooler does it for Fury/X, a ridiculously tiny open air blower is enough for TDP limited Fiji (clock GCN to >1000MHz and you have a problem, nothing new here). Maybe GM200 could do well enough if TDP limited with such a small heatsink, but then you probably can't fit it in such a small PCB, not without some witchcraft. Maybe higher density GDDR5 modules could help with that.
Considering that nV is doing a node shrink (harder than usual, planar -> finfets) + new architecture, using something similar to GDDR5 would eliminate the jump to a new memory type and allow more room for success (we all remember Fermi and the 40nm+new arch+GDDR5 combo...). Still, I don't expect them to use GDDR5X for GP100. The high end requires HBM2! They have a 250-300w TDP budget for the entire card, use a power hungry memory type (GDDR5X) and you have less and less of that available to the GPU itself. Clock HBM/HBM2 high enough to reach, say, 1TB/s and you still won't be nowhere near the power required to drive GDDR5/X to half of that bandwidth on a 384 bit bus. Yeah yeah, node shrink + new arch is a huge deal in terms of performance brought to the table over actual 28nm hardware, but why limit the increase with an inferior technology? Especially when the absolute performance and perf/w enabled by these other two new factors is even higher than before, every watt counts more than ever!
Fiji allowed GCN to become what it is today (even after almost 5 years) because of the improved power management and the freed TDP from the use of HBM that could now be allocated for the GPU itself. nV stands to benefit from this, too.
GP100 ought to make the jump to HBM2. Hell, GP104 could also be a monster with HBM2 (just see them price it like the 680/980, margins should be enough to justify the change). If Pascal supports both memory standards, then, well, nV is the master of marketing and apple-like brainwashing, they could very well sell a GDDR5X flagship and then make the change to HBM2 once it's 100% ready. Going forward, just leave GDDR5/X for the low end, that's where it belongs.
Just no more 970-like shenanigans. That's all I ask for this next round.
--------------
Question. Is the Fury Nano at all diminished from it's other Fury brethren? Clocks? Shaders? ROPs? Texture units?
Nano has a full Fiji, capped to 1000MHz / 175w maximum, and configured as such that
powertune throttles the clock speed/voltage according to the load.
The lighter the load, the less power consumed, the more the average clockspeed is to the 1000MHz maximum.
If you lift the
powertune setting from 0% (175w) you get an almost fixed 1000MHz clockspeed on most games with a little more fan noise and temperatures, although this depends too much on the workload at hand. There isn't much headroom because the power delivery is cut down and targeted for that TDP. Considering how small the heatsink is, it manages quite well. If you overclock on top of that, well, it gets overwhelmed. Nothing unexpected here. GCN is most efficient at the 800-1000MHz range. Once you get higher than that, power starts to get out of control (at least for Hawaii and Fiji, Tahiti/Pitcairn aren't as dense and don't "suffer" as much). They should try to fix this for Arctic Islands. Maxwell performs so well in part because of how well it clocks.
For how unbalanced GCN is in its Fiji form relative to what Hawaii has to offer and how much hardware it packs for non gaming workloads, it's nothing short of amazing for an almost 5 year old architecture. AMD wasn't kidding when they stated GCN was in for a long time when they presented it with the 7970. DX12 eliminating AMD's subpar DX11 driver (although good enough) and overhead should make (and is already making) things more interesting, but that's something for another time and another discussion.