2900 cf scores over 20000 in 3DMark06 on air

thats nice but useless i wonder if the 2950xt will perform better with its redesigned core in real world applications.

well it should be renamed since its a re-designed core and 65nm process. Hopefully they would have UVD so it would actually have HW/SW HD acceleration.

Well you see it doesn't translate well, which is why I hate raw Bungholiomarks for performance previews/descriptions.

It's good to find out max core/memory overclocks and for testing driver changes and system stability, but it doesn't predict/translate to game performance.

The biggest problem is these new designs can text the stress tests and skew them exponentially where before it was a 16/8 - P/V design now it would be like 24 P&V(&G) and you can dedicate all resources to tests meant originally to stress specific parts of a more spread out design. So some tests that were now harder can now get 22 vertex engines and 2 pixel shaders (like in 3Dmark05's Canyon flight) and act in a way that few games would call for. Also the texture loads are also not similar to a typical game.

So such benchmarks are no longer good at predicting performance since it's a balance of the sum of all the parts not just the extreme of each stage individually, and that's where the R600 and GF80 run into problems when whatever is called for is not something that that do well outside of dedicating all their resources to it.

So like the wonderful mid-range numbers in 3Dmark they over estimate game performance.

The 3Dmark world records are interesting to what they tell us about the physical limits of the chip and boards, but doesn't offer much insight into gaming.

Nah, Rydermark is interesting, although ugly IMO, but no single benchmark is a good estimate of 'games' just of the power of components that may make up a game.

The only good predictors would be D3 performance is a good predictor for Quake4 performance, and HL2 good for CS:S, etc. And even that doesn't always translate well.

IMO, the more benchies/info the better, but nothing does it all, and most people don't even look at the most important parts of the benchies they have like 3Dmark, and instead just focus on final scores not the more important composite parts that give insight into the architecture.

If you look at the decoding, the only thing that goes to the cpu is VC1, which if using it, has a small effect on your cpu. I wish the 2900 has full HD decode or UVD as well, but the loss isnt that great.

Oh, please forgive me

UVD isn't there and few people other than the AIBs said it was, however AVIVO HD feature and UVD itself are not the exact same thing, but they are related. The features of AVIVO HD will be done by dedicated hardware on the HD2400 and HD2600, while the R600 will have to use the shaders to fill in the gap in dedicated hardware support. This is detailed pretty clearly in both articles you link to below. The difference is in order to do it on the R600 you need to make it do it with the programing of the drivers and software like the articles also mention.

http://techreport.com/onearticle.x/12552
http://www.dvhardware.net/article19343.html

So while the R600 doesn't support AVIVO-HD yet, because they haven't activated support (BTW does everyone forget the original AVIVO and PureVideo lack of support right out of the box, until they added software to enable all the features?). The latest betas are eginning to add support but it's still early sofar with a few glitches from what I've read.

Like the Abyss points out, it's tempest in a teapot since it's features not available on any other card right now anyways.

Main thing is right now there's only a few differences from all these and even the previous generation in that the new cards can accelerate the BlueRay and HD-DVD DRM decryption. So if you play protected disks they are faster on the new solutions, but play downloaded H.264 content and you're about as fast with a GF7600 or X1600 as with the GF8600/HD2900/GF8800.