I'd like to see the article you're getting this from; also awhile back I said Nvidia could use OpenEXR HDR with Super Sampling antialiasing but everyone though I was an idiot. :lol:
I remember that thread, but I doubt it was I or Cleeve who would've thought you were an idiot, just wanting to know how it was done. I remember we discussed it 8 ways from Sunday! I had a feeling that it was done in software, because you could do that as explained above, but it requires more coding from the developer, moure CPU resources, and it's much slower than native (although may be faster than 2 loops required to do it in nV's GPU hardware). Man I tell you that original discussion set me off, on a very difficult journey to try and find out WHY it was possible, and what the owrk around was, and what the physical difference were that made it possible for ATi but not nV. And it's still not truely possible for FP16 throughout in hardware, but this would explain how they do it in software and I would imagine this would allow it to be FP16 throughout as well.
As for where I got it, you want me to reveal my secrets, after I spent so much time searching after that other thread!
It was in a recent GF7950GX2 review I saw, I'll pull it up again when I get home from work.
I'm also unsure how accurate the performance stated is; with Nvidia's latest drivers I've gotten a solid 10FPS more per second on Age of Empires III on 1680x1050 with minimized UI enabled. In general super sampling should actually be higher quality than multisampling when comparing the same level;
But it's NOT the same level, re-read their statement it's 2X (2.25) SSAA (as determined by software) versus ATi's MSAA levels (in this case they compared 4XMSAA to the software's 2XSSAA, and they said the hardware 4XMSAA > software 2XSSAA. Sure if it was hardware 2XSSAA versus 2XMSAA then yes you'd expect it to be better. But considering they are limited to the implementation it's not surprising. Heck ATi could probably do 6XMSAA with better performance, and it would be a factor better still, but they likely thoguht this was the closest comparison between the two (I assume 2XMSAA on the ATi would be faster still).
when using it my performance on games like F.E.A.R. is not much worse than Multisampling, if any noticeable difference.
But remember it's a different implementation, and it's also involving HDR, and it doesn't mention if the HDR is done before or after the SSAA increase and then filtering. The filtering process would involve the use of edge half-tone detection that might mess with the HDR implementation. something a little more detailed would be great.
I would like to think this could be attributed to Nvidia's unused pixel processing power which could help to keep super sampling moving, although this may not really be the case.
Except for the fact that in this case the SSAA would be at the software level done by the host, not the GPU, so nV's unused processing power would not be available. And since it's HDR that is another performance hit, you likely don't have much unused performance left in the ROPs which is where the SSAA would be done in hardware. Like I said the only two options are Software which is the way it is done in AOE3 or Hardware using 2 full loops through the ROPs using int8 targets for the second loop.
In effect the software method AOE3 uses might offer better IQ than the 2 loop method at the same setting (but it would let you choose 4XSSAA, unlike the software method), but both take a performance hit, the question is just how much each takes, and whether the IQ differences are enough to matter.
EDITed broken quote.