Very interesting and in-depth article on the NV40

[jimmyjames123]

Very interesting and in-depth article on the NV40:

http://www.3dcenter.org/artikel/nv40_technik/index_e.php

One of the interesting points made in this article:

--Long pixelshaders heavily using arithmetic calculations are shader-bound, though. Hence, considering NV40 was made for future games, we think there is no need to bother about NV40's bandwidth-issue today.


They also talked about NVDA's new AF implementation, which apparently is still slightly better than ATI's current implementation with respect to image quality, but of course is worse than NVDA's current AF implementation with respect to image quality. They noted that AA is very similar to ATI's current implementation, but ATI still has a slight edge. They talked about whether or not we need PS 3.0, and said that "it's always good to have options". They also mentioned that developers will enjoy and appreciate some of the features on the card. Finally, they mentioned that this 6 series architecture was designed to effectively scale, unlike the previous generations of NVDA hardware. This sounds promising for those who are looking for budget cards with trickle down technology.

Hopefully, there will be a driver option that allows one to enable the AF-style that was used on the previous GeForce chips, so that users have the option of increasing anisotropic filtering quality further if desired. Also, hopefully NVDA will expose some of the hidden AA modes, like the 8xAA mode with 4xMS and 2xSS. That should remove most of the small and few criticisms revolving around this card.

 

[jolle]

interesting stuff..
easy to get excited tho hehe, have to keep in mind alot depends on:
Feature implementation in software (like FP16 textures)
Real world performance..

But nice read tho, looking forward for the cards to hit the market and get reviewed for real, and compared to ATIs part..

 

[Sandwich]

nothing new then.

 

[WaltC]

One of the interesting points made in this article:

Long pixelshaders heavily using arithmetic calculations--are shader-bound, though. Hence, considering NV40 was made for future games, we think there is no need to bother about NV40's bandwidth-issue today.

Pixel shaders are certainly bound by their own performance, as opposed to local bandwidth considerations. Longer pixel-shader instruction chains, however, will still render slower than shorter instruction chains, and the slowdown is fairly predictable based on the length of the instruction chain--the longer and more complex--the slower the render in terms of fps results. IE, you trade one bottleneck for another--you exchange the local bandwidth bottleneck for the bottleneck imposed by the shader. A long/complex enough ps instruction chain could easily result in fps performance far beneath the limits imposed by the local bus bandwidth otherwise. IE, "long pixelshaders heavily using arithmetic calculations," is no defense against subpar frame-rate performance, and quite easily may cause such performance slowdowns in 3d games. So, you are probably going to want to limit your chains to keep them as short, and as efficient, as possible to obtain your rendering IQ goals with satisfactory performance.