That would not be surprising, because GK110 in its current form would really offset any of the current efficiency benefits of the GK104, not to mention that it is mostly tailored to compute performance.
That is just plain wrong.
If you look at K20X, you have 20% higher TDP than GTX680 (235W/195W) but should have about 20-25% higher 3D performance (due to higher compute power and bandwidth). Additionally, you have 3x the memory which of course also consumes power. 1GB of fast GDDR5 consumes around 8-12W depending on current operations.
Second, "compute" performance IS gaming performance (if sufficient bandwidth is provided). More accurately, SP compute performance. What you probably mean is DP compute. But a GK110 SMX has the same amount of TMUs and ALUs as a GK104 SMX. There is no reason at all that it would perform worse. Additionally, GK110 has more cache, which should help it to utilize bandwidth a bit more efficiently.
The improved DP compute capabilities and HPC features cost transistors and make the GPU significantly bigger. That's it. Increased power consumption/clock in gaming comes mostly (if not completely) from the increased number of units and the wider memory bus.