Originally Posted by TheRockMonsi
I believe he's aware of that; from how I understood it, he was saying that why would NVIDIA go through the hassle of making a relatively big card when the smaller one performs very close to the same (questioning the business decision behind making the bigger card).
Well, he explained it perfectly. If the thermal envelope of the GPU is low enough you don't need a large polyphase energy supplying circuit with an enormous array of VRMs. My guess is that they included this as a premium on the 680 for overclocking the card. If they can cut costs on the 670, they will do it.
People may disagree on this point with me but from a design perspective, GK104 is pretty much a midrange GPU even though it outperforms all currently released GPUs. You could see the GTX 680 as a beefed up GTX 660Ti with power circuitry to enhance the achievement of higher clock targets, and the GTX 670 is like a standard GTX 660 which does not have this premium. This SKU based on a crippled GK104 was not designed to take the performance crown after all.
The slight performance hit from removing an SMX out of the ASIC may be caused by either inefficient thread scheduling in a full GK104 (logic units in the pipeline aren't saturated with data). Or more obviously is that GK104 is being bottlenecked by it's memory bus since it's not wide enough.
Either way the GTX 470 and GTX 480 did not differ that much in performance either clock for clock while having 1 SM and 1 memory controller disabled (same as a GTX 470 and GTX 570).