Finally a proper explain.
8+8 card can pull more than just 375W, as proved with 7990. Just make sure your PSU is single rail, or each 8 pin is on a separated healthy 30A rail.
By the way, 780ti is a 6+8 card, and it can pull a lot more than just 300W.
Right. There are actually many cards that do it. Before the days of
software to limit tdp so amd/nvidia could upsell you powertune and the like many volt-modded 4850's way out of spec. The same game was played in reverse when nvidia's 500 line essentially was clocked to the balls end of a pci-e spec for their plugs and overclocking brought them substantially over. The fact of the matter is while you could say pci-e plugs are an evolution of the guidelines of the old 'molex' connector, which is all well and good, they are over-engineered by a factor of around 2-3. This card just seems to bring it down to around 2.
It can, but varies on motherboards design. A prime example would be the recent GTX 750/750Ti reviews that used the reference (no aux power) design for testing. More than a few results (esp those which overclocked) yielded a power draw in excess of 75 watts.
Very interesting! Thanks for this. So it's a factor of around 2 then, if not the card simply limited to that power draw (to stay in the 75+75w spec) and it's even higher.
Well then, I guess you could take 936 + at least 141w then. When you factor in the vrm rated at (taking his word for it) 1125w, it gives an idea of what the card was built to withstand (which is insane). It seems engineered for at least 2x spec, which sounds well within reason for anything anybody is realistically going to be able to do with it. I doubt many people have a power supply that could even pull that with a (probably overclocked) system under load, even with an ideal cooling solution.
Waiting for the good ol' XS/coolaler gentleman to hook one up to a bare-bones cherry-picked system and prove me wrong (while busting some records). That seems pretty much what it was built to do.
On a weird sidenote, their clocks kind of reveal something interesting inherent to the design. First, 1018 is probably where 5ghz runs out of bandwidth to feed the thing. Second, 1018 seems like where 28nm would run at 1.05v, as it's in tune with binning of past products (like 1.218 for the 1200mhz 7870 or 1.175 for the 1150mhz 7970). It's surely a common voltage/power scaling threshold on all processes, but in this case half-way between the .9v spec and 1.2v where 28nm scaling seems to typically end. Surely they aren't running them at 1.05v and using 1.35v 5ghz ram, but it's interesting that they *could* and probably conserve a bunch of power, if not even dropping down to .9v and slower memory speed. If I were them I would bring back the UBER AWSUM RAD TOOBULAR switch that toggled between 1.05/1.35v for said clocks and 1.2v/1.5v-1.55v (because 1.35v 5ghz ram is 7ghz ram binned from hynix/samsung at that spec). That would actually be quite useful for both the typical person that spends $1000 on a videocard (like we all do from time to time) and those that want the true most out of it.
I've got a long way to go before I reach the lack of understanding and trolling you Xzibit in the Nvidia threads 
