Well, what
do we know about the R420?
First, let's look at the history.
It was stated by some people in the know on this board, that the R400 was cancelled due to it being too good compared to the possible competeition and being a bit too ambitious at that moment. Which makes sense, if they gradually increase the chips, they can sell more then when they just launch one that is so good that the games need years to catch up. In the mean time, the production of gate arrays and memory chips can catch up and make that chip more interesting commercially.
The R400 project was merged with the R500 project, while some of the developments were combined with the R3xx to make the R420. It was stated later (and confirmed by the deal with Microsoft for the XboX2) that it will be a DX10 chip (with PS/VS 4.0).
If we compare the actual performance of a quad pipe from a 9800XT with a quad pipe of a 6800, we see that the 6800 is marginally faster in full-precision (1..15%, depending on your calculation, 5% being the mean value).
nVidia has switched to the ATi way: a lot of lean pipes. But who is to say that ATi didn't switch a bit to the nVidia way and put more functions in their pipes? For example, if they upgraded their current design to 2 full ALU's, 2 texture units and two mini-ALU's per pipe, they DOUBLE the throughput of each pipe in just about any real-life situation.
But the question we should ask ourselves is: as the R500 is DX10, VS/PS 4.0, what part of that design did they add to the R3xx? By the existence of the 2.X spec in the DX SDK, we all assume that that is the spec that will be used by the R4xx chips. And ATi has just about confirmed that the R4xx is not going to be PS/VS 3.0.
It would be a supreme joke when the high-end parts are actually something like PS/VS 4.0 instead
Dynamic branching is more or less confirmed to be no part of the R4xx chip. Nobody has as much hinted that the vertex shaders are not 3.0 or 4.0. Which would make sense: it takes far less transistors (as there are less VS to begin with), while it would give the largest gain.
So, how does a R4xx with VS 3.0/4.0 and PS 2.X sound, probably with only static branching and/or BREAK in the VS? With expanded functional units in the PS, while keeping it at 24 bit?
That would make sense and be completely in line with all the hints dropped.
