First you post a link which contradicts your statement. (3x37%=111% overall or 3x41%=123% in low resolution).
Overall at TPU is a useless metric. Who buys a $600 GPU and games at 1600x900?
1600P
Bonaire = 31%
780 = 94%
MSI 780 = 100%
3x Bonaire won't beat after-market 780s in demanding games/high resolution gaming.
Then you somehow implying that boinare is DP crippled like kepler...
It has nothing to do with Kepler. Bonaire is DP crippled. There is nothing to imply. It's a fact. It has 1/16th of SP.
http://www.anandtech.com/show/6837/...w-feat-sapphire-the-first-desktop-sea-islands
If you just take 3x Bonaire chips, you have a DP crippled flagship. Once you add/enable those double precision transistors, do you think your die size/power consumption won't be adversely affected?
7790 takes 69 watt in average conditions (techpowerup review you linked) and 103 watts max.
Average power consumption doesn't tell us what the card peaks at in games. This is why I always link to Peak power consumption. Why? Peak GPU power will happen in games. Even if it happens for 5 minutes in a very demanding scene, that stresses the power supply, VRMs and the reference cooling system. When AMD designs a GPU, they must look at the peak power consumption in games because if the card fails at peak load, it's not suitable for sale.
Bonaire peaks at 77W. Either way, you missed the point that double precision transistors are not free - not in the form of power consumption or die size. Therefore, even if you triple Bonaire's peak power you still have to account for DP hit, unless AMD will cripple its flagship.
After-market 780s are
way faster than a stock 7970GE. At this point with such small premium for after-market 780s, using a stock 780 as a point of reference that 9970 has to hit is not realistic either.
Finally, GTX680 was 10-12% slower than 7970GE but it accomplished that feat with 33% less memory bandwidth and 25% less shaders.
Titan has 48 ROPs, 224 TMUs and 2688 SPs. If you triple Bonaire, you are well under 224 TMUs. The one possibility that could come into play is that the major bottlenecks in Tahiti are resulting is less than optimal optimization of 2048 SPs. If that is true, then once the ACE/TMU/ROP bottlenecks are opened up, the SPs may be used more efficiently.
The longer 9970 is delayed, the less impressive it will be because every month we are that much closer to 20nm GPUs. Even if AMD undercuts 780 by $100, NV could just drop the price on after-market 780s to $550. AMD has lost the momentum after 7970GE and bringing out a card that barely beats 780 by 5% is not good enough.