Originally Posted by Ultracarpet
well if it did in fact scale so well, why on earth would they even go to 20nm? they would have AMD beat in every aspect by like 50% while being able to keep yields and margins good at 28nm. I don't know how big of a change Maxwell is architecturally, but it might affect the way it scales on 28nm with a bigger die and it's why they are needing to go to 20nm.
I'm not speaking with certainty here, really just playing devils advocate.
-We don't know what AMD is doing for their next GPU generation
-20nm has probably been planned for years, there's literally no reason not to do it because it would waste a ton of money spent on R&D already.
And playing the devil's advocate is fine, but to actually make the scaling argument you need some proof. Unless you're just purely speculating on some scenario.
The only real arguments in favor of worse scaling than kepler that I can think of are:
-GM107 doesn't have any GPC interconnects because there's only one GPC so adding more would require the extra logic and might make the scaling worse
-Since cache size is just purely decided by NV and isn't tied to a specific GPC, SMM, etc. count the cache amount could be cut down (per GPC) and this would increase the extra power consumption needed to access VRAM since it would have to be done more often.
Then again for good scaling you could make a ton of arguments:
-15% higher transistor density would pretty much automatically improve cuda core / ROP / memory controller / etc. count over 28nm Kepler designs, you don't even need architectural changes for this.
-One Kepler SMX is much bigger than one Maxwell SMM yet the Maxwell SMM offers 90% of the performance of the Kepler SMX (35% increase in per core performance)
-Maxwell clocks higher than similar size Kepler designs
-Cutting down the cache amount would give room for bunch of cuda cores / rops / etc.
-Even a small decrease in power would drastically improve high end (though unless maxwell scaling is incredibly poor the decrease in power consumption wouldn't be small) as current high end is pretty much completely power limited, so the power efficiency would actually help with performance more in the high end than low end.
All in all "because scaling" is a pretty bad argument.