Are you lumping compute in with tessellation ? Because in the games you've used with the largest differences; tessellation looks to be the deciding factor in the differences. You expect this to play out the same way with Tahiti vs Kepler ? I suppose that's possible if nvidia can reduce the performance impact tessellation inflicts on their cards more so than it does currently, but that sounds like a stab in the dark. You might want to re-read the
results of multi-threaded rendering for nvidia. Performance
doubled, that's a significant gain.
You may not want the one benchmark that helps to prop up your whole opinion more than the rest to be discarded, but when there is a result so supect and defying logic, it needs to be looked at closer. Have you played B:AC ? It looks like a console port and the graphics are not groundbreaking, contrast it to BF3, Metro 2033 or Crysis 2, games that show a reasonable performance disparity and visually are exponentially more impressive, not the ridiculous 100%. The 100% jump seen there begs suspicion in a game that is one of the least impressive graphically in that entire list. It's also a suspect result that most skews your entire summation.
Per Shogun 2, that is the one game you are using where it is not running on its maximum settings. It's also another game that cannot be run on a 1GB card at 1200p maxed out, another bench better left out because you took it from another source and it's not being run on its highest settings.
Again, I don't disagree with the 480 being a better card than the 5870, again - we knew this 2 years ago. But, the way you are trying to draw a comparison between the 480 and 5870 and use that to predict how kepler will compare to tahiti is not valid. Firstly, because someone could choose a list of DX11 games that gives a completely different summation than your arbritarily decided on list has. Secondly, because you are assuming performance characterists with no information available at all about how an unreleased card will perform.
If someone had used a comparison of the 3870 to the 8800GTX to come to a performance estimate on how the 4870 would compare to the GTX280 they would of been way out in the weeds. You've done more than compare GPU architectures here; you've compared VRAM buffers, driver differences, a suspect benchmark showing 100% gains and different testbeds and testiing methodologies here.
Appreciate your enthusiam and efforts but you've just reiterated the results of 2 year old benchmarks and added in a dose of speculation on an unreleased product with no proven foundation in the data you have here.
I get that there has been nothing new for a long while now to talk about with nvidia graphics cards. So we're going to back 2 years and reiterate on the 480 vs 5870 and try to make it relevant for a new architecture vs an unreleased and unknown architecture ? Really ? Hopefully
this does not pan out to be true or we'll be hearing more wild speculations like these for the next seven months until September.