Well, idle power consumption is a bit lower than the 5770 (or the 5750). Load power consumption is a bit higher, but not by much (comparable to the idle power consumption advantage). Noise is also lower.
Yes, I think nvidia deserves some credit for providing a decent reference cooler even with this class of cards (its only problem seems to be with SLI if two cards are packed densely, but that's not really a good idea anyway). Doesn't look though that too many manufacturers will use it.
However, it also doesn't perform as well, so its probably really seems to be price/performance, not power/performance.
I think the idle power advantage is due to downvolting memory - no matter how it is achieved, it's definitely nice to see things are improving there.
I think nvidia chose wisely not to try to beat HD5770 with the reference card by just upping clocks and voltages - everybody would have complained about a card with performance similar to HD5770 but power draw of HD5850 (or GTX460, for that matter). Still, the higher clocks (and voltage!) compared to GTX460 already show up, but the compromise isn't too bad.
Overall this looks all solid, beating GTS250 (with lower power draw) on average easily (OMG nvidia can finally retire g92b!!!), sometimes even close to GTX260, so this definitely is an improvement.
Looks like AMD is willing to fight back with price drops of HD5770, about time (well for consumers...), which could make life difficult for the GTS450 - though HD6xxx will probably be a much bigger problem.
One advantage GF106 (like GF104 too) has seems to be it's a bit less dependent on memory bandwidth compared to AMD's chips, hence memory chips should be a bit cheaper. It is also DP capable, which some people might miss from Juniper.
The die size still has me puzzled though - GF104 vs. Cypress is (depending on which of the two numbers you believe) either same die size or 10% more, yet GF106 vs. Juniper is nearly 50% more, despite both chips being mostly half of their respective brethren. In theory GF106 should downscale more effectively (due to less shared logic which doesn't downscale with Juniper) - granted Juniper also dropped DP which should save some transistors but still the difference is so big that even the (not even used) additional ROP partition imho isn't enough to explain it. At least the larger the chip the more easy it is to cool but that's certainly not the reason for it to be so big...
