I guess though power draw isn't actually lower than that of a gtx 770 after all despite the lower TDP. Still, it's quite an improvement on perf/w. tomshardware has some very interesting power draw numbers, apparently with the right load these cards can be stressed to a level which is easily beyond the rated pcie power connectors even (50% above TDP, right at the level of r9 290). I guess that puts some things into perspective.
Any device in this class can, because it's not a challenge for these chips to physically draw way more power than what they are rated for. It's not particularly challenging for them to draw enough power to damage themselves. The real challenge is finding ways to not let chips kill themselves.
I don't dispute the likelihood that Maxwell has improved its voltage and clock management at least somewhat, but if its DVFS is improved, I have missed whereTomshardware's numbers measure clock and voltage levels.
Using a scope to give a jaggy power graph doesn't contribute anything but noise, because of course complex power supplies for complex chips running complex workloads don't draw a flat line.
If they want to claim Maxwell is doing the same thing as Hawaii, they should use their new setup on both and log their clock/voltage steps before stating it.
Power spikes happen at extremely short intervals with or without enhanced DVFS, and if there are particularly high spikes it can be a sign that the solution is actually too slow to catch abrupt ramps in demand.
If the higher than TDP power numbers for their compute load are sustained, then that's definitely not a sign of a significantly improved DVFS.
Spikes above TDP are actually fine within the constraint that they not take more than some thermally significant period where the thermal solution is expected to absorb transients. Beyond that, more advanced turbo methods can exceed it for longer if they've modeled the cooler's thermal capacity and realize there's slack.
If it can be forced into a sustained draw above TDP, such as was given for the reference design tested at Tomshardware, that's what happens when the chip isn't doing anything particularly leading-edge in terms of DVFS.
If there were an area where I do think AMD does better, it would be in this realm, since it likely buys them several speed bins that would be lost due to guard-banding and a more conservative TDP.
However, like Mantle, such measures only let you get as far as the architectural laurels AMD has rested on for several years can take you.