Originally Posted by 00firebird
This is pretty much proof positive they would take back ever releasing the 670 if they could, as it offered buyers way too much value. the 660ti wasn't good enough and the 680 was too much. So of course they rebuilt the product lineup around removing the 670.
I'm not so sure that's "proof positive". Maybe enough proof for civil court, but certainly not for a criminal case.
Anyhow, I'm not sure what they could've done differently back in early/mid 2012 to fill out their lineup.
Just getting to 6008MHz memory was an accomplishment at the time (as the 580 was only 4008MHz), so goosing the 680's memory clock any further was out, and reducing the speed of the 670's memory would have been counterproductive. The ROPs were by design directly proportional to the number of 64-bit memory controllers (8 ROPs x 4 64-bit controllers for the 256-bit 670 and 680), so those couldn't be independently cut either.
All that leaves is the number of SMX clusters (which is directly proportional to CUDA cores and TMUs). Reducing the 670 to six SMXs would've been too severe a cut to justify the $399 price point (and would've had knock-on effects on the 660 Ti as it obviously couldn't have more due to marketing considerations), but having seven SMXs was too close in terms of performance to the 680's eight (even after reducing the base and boost clocks of the 670).
In short, NVIDIA did what they could with the 670, given the limitations on cut-downs that the Kepler architecture provided and the number of price points that they had to fill. Now that they're on a "refresh cycle", NVIDIA has more freedom to tweak the price/performance ratio (especially since they don't necessarily need to build an entire full line of 7xx cards).Edited by svenge - 6/21/13 at 3:24pm