Originally Posted by raghu78
You have to understand that the transistors used for Double Precision units (8 per SMX in GK104 vs 64 in GK110 per SMX) are not going to help gaming performance. The GK110 has an architecture which is purpose designed for excellent HPC and workstation performance. Features like ECC, Hyper-Q , Dynamic parallelism do affect die size and power. Perf/watt will take a hit. the desktop GTX 780 will have to work within a 250W TDP so that means lower clocks as can be seen in the Tesla K20 running at 705 Mhz..
While HPC does take die space, the additional shaders + ROPs and wider memory bus will also greatly affect gaming performance.
As far as HPC units affecting perf / watt, they propably don't, as they are likely gated down when not used. In this case die complexity will add more to that, as such a big die with low leakage in much harder to make than the smaller one.
As far as clock go. Kepler as is is either constrained by ROPs or memory bus, so that overclocking is not giving theoretical max increases in most cases, so again by eliminating bottlenecks can give a greater result in certain scenarios than pure clock speed itself.