Until Kepler that is exactly what happened the high end GPUs were near the upper end of what was possible without going completely silly and the cores that didn't make the grade were used in the rest of the lineup - high end and volume don't generally go together - as soon as you are talking volume you know you are talking a mid range product and not high end when it comes to GPUs. A "volume high end product" is not the definition of a high end GPU.
There is a huge gulf between the GP100 and the 1080 - easily enough space to make something that is more like high end in the region of the TX stats. 16nm is more expensive than some previous processes but not to the extent you are portraying and the process itself quickly hit a good level of maturity once it was out of risk - nVidia is pushing the clock speeds a bit more than the process was really envisioned for which has had some impact on availability.
If you look at ~300mm2 cores that were around about the same performance faster over the previous high end on the node before them we are talking 8800GT, GRS250, GTX460, etc. you have to completely rewrite the rules to make the 1080 high end. Meanwhile nVidia is laughing all the way to the bank off those who buy into it.
So you have again completely ignored the possibility that the technical problems of the more recent 'nodes' may have changed the economics of GPU's so that releasing bigger dies up front might not make the best sense. Your 'high end' metric has already been rubbished there are other limiting factors then net power draw for a GPU. Net power draw may have been more of a limiting issue for NVIDIA at and before Fermi's release not so much now.
Up until Kepler GPU's were to an extent the limit of what was feasible due to their power draw its just that the limiting factors have most likely shifted since then.
Why is it so hard to understand that something being technically feasible does not mean it would make sense economically?
The ever reducing quantity of companies (fabs) making computer components suggests that the ongoing costs of successive node shrinks are becoming prohibitive
http://www.extremetech.com/computin...its-of-smaller-processes-and-new-foundry-tech
Last edited: