There is so much inaccurate info going around.
This is a post based on some pretty technical info, and if you want to understand its basis, I can recommend a place like Beyond3D to you to begin. But, for now, I'll just cover an outline that comes to mind to address some of the things I've seen said here, and some other places.
Let us split a card into two aspects temporarily:
Processing (the core design and clock speed) and
Bandwidth (the memory speed and bus width).
Processing
Let us split this topic further into two, because it is necessary to explain why so much of the info going around is bad (this includes some popular sites that "sound" technical in their discussions, but aren't being accurate).
Pixel processing, or "what is done to draw detailed colors and object surface effects".
Let me explain the central issue with the current generation of cards that nVidia spends a lot of time trying to hide: In pixel processing performance, the Radeon cards simply dominate completely, top to bottom, corresponding FX cards. nVidia puts out competitive performance results by (slightly simplified): 1) (cheating) working hard to to remove other performance penalties from the picture (like not offering trilinear filtering in Direct3D when asked, even though the card can offer it), 2) (cheating) replacing shaders with other shaders that actually do less work, 3) (optimizing) using a shader optimizer that can produce better performance with the same work, just being done more efficiently.
(suffering succotash...I hope you didn't read that out loud, or you'll have to wipe off your screen.
)
Now, 3) is very good, and ATI is working on this too, but there is a problem here from nVidia, which is probably related to why ATI doesn't talk too much in marketing about their doing the same thing as 3). The problem is that nVidia calls
everything "optimizing", and has for a while, even though only the last step (which has only fairly recently been evident for nVidia) deserves that name. This confuses conversation about optimization, because people think the "optimizer" is doing amazing things, when the person could do the same thing themselves by turning off trilinear filtering or selecting inferior shader effects for the games that offer it.
Explaining this requires very technical and detailed info, and careful examination and discussion...video cards are not simple things. Fooling someone about it simply requires witholding that info. This is why many people view the above as a "rumor" instead of something already proven if you know where to look and can understand what is being discussed.
...
Processing performance parity or superiority for nVidia cards is almost completely an illusion (i.e., lie). In "DirectX 8" applications, in some specific shader workloads, this might not be a lie when comparing a 9700 to a 5900 (or 5800, which executes DX 8 at similar speeds to later cards), but with the higher clock speeds of the 9800 products, with the 5800/5900 products not increasing clock speeds, figures that continue to show this are pretty much always due to other factors (the FX cards have some advantages, and there is the cheating as well).
The exception is the drastic price drops on the 5900 SE (aka XT
) cards, which are actually a fairly high end card priced lower because nVidia is fooling less people that count (card manufacturers and system builders like Dell), because they do factually compare fairly well with the 9600 cards in processing performance: they have a higher processing performance potential, but also have some serious performance issues that whittle away at it for medium to long shaders of the moment. Given its target market, it could very often be a clear leader here, if you could get rid of the cheats listed above compromising on image quality (which widen the performance leadership by deception, but produces typically inferior output).
Also, in processing
capabilities, there is the specific case of the 5200 has more capabilities than the 9200, but the problem is the 5200 is incredibly slow at capitalizing on the difference. A large portion of its processing power is even done at less precision/range than the 8500-9200 (12 bit fixed point versus 16 bit). The 8500-9200 actually have significantly more processing power for the fairly flexible shader functionality they offer.
Finally, you might have heard that the NV3x has more advanced shaders than the R3xx, but, for even longer technical reasons than the above, that isn't
quite true unless you look at just the list of instructions alone (which is misleading, since it isn't just instructions that matter). The R3xx offers a set of features to go with its shading instructions that allow it to quickly and efficiently offer advanced capabilities that the NV3x chokes on.
...
Sorry for the length, but nVidia's year long campaign to hide the above makes it a complicated subject. For the rest, I'm going to cover simpler things (things where nVidia doesn't have such severe problems, so they haven't had reason to complicate things with so much misleading information).
Vertex processing, or "what is done to make object shapes and shadows"
nVidia compares a lot more favorably here. They do still suffer in comparison for some more complicated vertex processing, but they also often (without having to resort to drastic lying) excel at more simple vertex processing. Given a mix of these in games, this will tend to (in reality, not just the fiction current nVidia drivers try to establish) go back and forth in advantage between the card families. However, it is pixel processing that is the most performance limiting factor in games moving forward at the moment, so this is primarily manifesting in games that have already been published, and not upcoming games. Also, the high end Radeon cards have been climbing in clock speed, improving their standing in comparison, while the high end FX cards have not.
Anyways, Doom 3 might have been like this if published last year, retaining a slight edge for (some) existing nVidia FX cards to their comeptition, but given that some things Carmack has talked about experimenting upon for Doom 3 having already been implemented in some other games, I suspect that Doom 3 is evolving a bit during its delay to incorporate more of the advanced effects he reported the Radeons as performing more quickly.
Bandwidth
nVidia compares more than favorably here...they have oodles of bandwidth (typically), and have a significant advantage over ATI. It is simply due to their significant problems in (back to this again) pixel processing, as well as some other factors that I didn't get to (like AF and AA quality and performance comparisons), that this advantage does them so little good for how much of an advantage they often have in competing products.
Again, the 5900"SE/XT" stands out, because they have a massive advantage here over 9600 cards (as long as they retain the 256-bit bus for a card in this price range). Actually, in combination with the Processing power it has, it ends up being, in general, a better card than the 9600 (aside from AA quality and power/size/cooling limitations), if only all of its capabilities could be exposed instead of being replaced with performance "boosting" image quality reductions. The only thing wrong with it (besides drivers preventing access to all capabilities, as long as they keep the cheats in) is that the "XT" name seems like such a premeditated low blow (which, however, doesn't change the product's strengths at all), and that the 9600 cards
can actually sometimes beat it in performance...with about half the transistor count and bandwidth
.
Of course, if this causes ATI to offer 9800 NPs at the same price, or you can find a 9700 card, it doesn't fare so well. This is why it is unfortunate for such future competition driven products that nVidia is heading down the road they seem to be in terms of their products.
...
But, anyways, for what you
asked, the 9700 should almost always offer significantly better performance and/or image quality. I hope I've explained why there is some confusion on the matter, if you found the above understandable.
Or wait a little bit for the 9800 Pro to drop in prices!!
