Evildeus said:
Name FX 5900
3 cards 299-499$
2 introduced in june one in july
GPU 450 MHz
Memory 425 MHz
256 bits bus
A bandwidth advantage, and looking like a large one...I think this will be the secret to benchmark selection. "256-bit is overkill, until we do it too".
I expect the $499 cards to be the benchmarks we'll see.
128/256 MB of DDR/DDR-2
AA-AF ameliorations ( :? ) INTELLISAMPLE HTC
I'm reminded of some of the hacked up RAM usage calculations I made before, for when the post processing AA might be used to offer performance increases. I think if this will be a possible use of the 256MB cards, and I'm also curious about changes for the 256 MB "mid range performance level" cards and their AA performance hit, maybe exposed now, or maybe depending on the Det FX drivers or some other later driver change.
CineFX 2 (
)
New shadow accelerator: UltraShadow (for DOOM 3
)
Very Xabre-esque (can I steal that term?)...not because of the names, but because of the names instead of actual details. If this would change after launch, it would be different, but it doesn't seem likely that it will.
2* shaders operations (than NV30 )
I think this is good news for PS 2.0 compliant performance, and maybe Direct X feature exposure in the drivers, compared to the NV30.
According to Nvidia 20-80% faster than a 9800 pro @ 16*12 + 4*AF/8*AF
This would fit trading extravagant RAM usage for performance for AA using the post processor, I think. Just a whacky theory for now, though.
But then the caveats would be that it was when you aren't using PS 2.0, have a 256 MB card and have enough room for your texture demands, and aren't using the equivalent of Application AF.
...
Further on the whacky thinking...if some aspect of AF quality depends on post processing, the lack of RAM might have limited the "new" AF quality for the NV30 as well...so you'd need a 256 MB card (and 4x AA activated) to get good quality AF with good relative performance) in a game with the texture usage level of recent games. This theory might be even whackier, but I think it fits something that could be spinned into "debugging", and maybe nVidia's nv30 not being as screwed up as it appears to be in this regard (as long as there is enough RAM).
Did 3dfx have any such aniso-like shortcut in their post processor usage? For AF being on, it would be sort of an adaptive texture supersampling but with AF sampling criteria, and I think a dependence on AA usage to restore texture sampling quality might be the reason Aggressive AA looks so horrible, and Balanced doesn't look measure up to Application (i.e., it might with 4x AA implemented in the post processor, which it couldn't be for the nv30 because of expensive and high clock speed demands needed for RAM, making 256MB impractical).
The only problem with this would be the price you have to pay in RAM usage which might preclude getting extra details from higher resolutions and larger textures (EDIT: Oh, yeah, and killing color compression effectiveness...you'd be doing supersampling and hiding performance loss in reduced AF performance hit and skipping a blend step, and this would make sense with a more complicated color compression scheme that prevented color compression from disappearing completely). Quake III or something with similar demands as a showcase application for benchmarking would fit this theory.
Oh, and [/rampant speculation]
