I think the answer is "not knowingly". Optimizations tend to be done for an
entire generation (or a few generations), not for specific chips. 

That said, if a specific chip is observed to be an outlier (performance quite a
bit lower or higher than expected) then that specific discrepancy will probably
get some attention.

I don't understand your Mafia2 numbers - the Linux version wouldn't be running
under Nine, would it ? If the Linux version were running under OpenGL, then
this would look like a fairly typical "DX renderer in the app gets a lot more
work than the GL renderer" scenario.