link it I was under the impression gddr5 uses less power per clock then gddr3 and it is considerably less.
GDDR5 and GDDR3 aren't meant to run at the same frequency. What would ahppen is if you try and hit a certain bandwidth, you could use a N wide GDDR3 interface or a N/2 GDDR5 interface. That saves you power, die area, etc.
The problem is that NV isn't keeping bandwidth the same, they want to increase bandwidth and get better performance. That costs more power, especially if they have a wider memory interface than ATI (and it seems like that is quite likely).
really that was very profound, performance still drops with ATi's tesselator, its around 30% performance hit when going to heavy scenes. Also stream out performance is very important as well.
https://a248.e.akamai.net/f/674/920...sselationForDetailedAnimatedCrowds_SLIDES.pdf
The point of tessellation is to substantially increase scene quality. If you just use tessellation naively, it does reduce performance.
The point was that if you use it correctly, you can get 2-3 orders of magnitude more polygons, and you reduce memory bandwidth. It does cost you, as you say, 30% performance. But ALU performance is cheap. Bandwidth is not and 3-4 orders of magnitude more polygons can be noticeable in some cases I would imagine.
DK