Generally, if you raise the clock and power-consumption bumps up to the 110% limit, then the GPU automatically throttles back.
If you are not thermal or power limited, this shouldn't be happening. What I found out with my GTX 670, that at lower temps, I had the maximum boost (1254 Mhz to be precise) rate maintained 99% of the time, sure my power limit was raised to 117%. Pretty sure, with a bios mod, you should be able to raise the power limit beyond the 110%.
I'd think the trick is to find the clock-speed where power-consumption just tops out below that 110% limit. If voltage can actually be lowered, then there are a few extra clock cycles to gain.
This makes sense, trial and error is what I say, and have a spare card nearby, should the flashing fail.
I'm fairly new to this -- like I said, I'd dabbled with it back in my GTS 8800 days. Didn't fiddle with the GTX 570 card, or the 780. I'm still wondering if I lost my mind for not looking around more for a used or stray 780, but I was star-struck by the news about Maxwell.
Fermi was fairly easy to tame though, no turbo boost, you could change voltage on the fly. Well, the big Kepler is more like 290/290x power consumption wise. Maxwell is much more efficient in using available resources, not a big loss for you, I think.
Over all that time, the software made it easier to overclock. It seems more like a marketing game. They make the card to almost guarantee that it will overclock so far within the power-consumption limit, so any noob who does his homework should find the setting at 1500Mhz core and 8000Mhz memory easy without touching voltage (if indeed it can be touched, but I explained that.)
Of course, they could increase the stock clocks instead, would have been better, especially for the people, that are not comfortable with manual overclocking. I suspect the reason was simple, though. Power consumption. They just found the sweet spot and decided to call it a day. And whoever needs more, they can OC.
Still, the reviews concluded that these were "very overclockable" to that limit. Of course the reviews were stellar, but the latest gripe over the 4GB shows a lot of things, no less about the complaints and customer perception than about "marketing people." And I still think two of these cards in SLI is a good bang-for-buck compared to two 980's
It's still a well-balanced card, if you can live with its obvious shortcomings.
Maxwell did live up to the promise. I figure with a stock max power-consumption spec of 145W, the OC puts it close to 160. I think you can run two of these bad boys with a 550W PSU with plenty of room to spare for "extra gadgets" along the way.
In my experience, the official TDP numbers don't mean a lot. There are similar cards with the same 970 silicon that would consume different amounts of energy, up to ~50w difference, easily. It's not just voltage, clocks, silicon. Its everything including the power circuitry/PCB design, that affects the total power consumption. Obviously, the reference 970/980 would be the most efficient and easiest on the power bill (out of the box).
I would firmly say, that the reference 970 (sold in Bestbuy) and the reference 980 are the best for 2-4 card configs. From performance-per-watt perspective as well. With down-volting / down-clocking, you could tweak any card to find perfect balance of "speed and acceleration", though. Just takes time.