Rollo, IMO a big purported benefit of DX9 is that anyone can use it to create nice effects, not just the Carmacks and Sweeneys of the world. TR:AoD seems like a good example of that, and it shows that Radeons may perform better in "unoptimized" DX9 games.
You haven't seen TRAoD running on a R3x0 board have you? The game is poor in every area, including visuals. If anything it gives people the impression that people other then Tim and Carmack
can't make killer looking games no matter what they have at their disposal.
Ben, sure Halo may be a decent engine to benchmark, but the default flyby benchmark doesn't seem to highlight any differences in PS2.0 performance.
The issue here is that people were expecting the rift in synthetics to be anywhere close to real world in games, that was never going to happen within the lifetimes of the boards, I'll expand on this below.
Carmack's thoughts are spot on ... for nV parts. The 9700P appeared to be the first card that could offer "next-gen" features with very playable speeds, and benchmarks seem to bear this out, no? (Or did the G400M and GF offer 60+fps with their bleeding-edge features? I honestly don't know, I wasn't as involved in the 3D discussion back then.)
This ties in with the last part. Back during the launch of the original Radeon there was a lot of talk about how much weaker its hard T&L unit was compared to the GeForce- that was true from a synthetic point of view. Real world either you supported the feature or not, no developer was going to rely on it overly much as the overwhelming majority of their audience was going to be lacking hard T&L. My comments from that era are still in the forums and people can feel free to verify the fact that despite my very strong support for hard T&L, I did state numerous times that the Radeon's considerably slower performance in synthetics would never materialize in games, devs would not allow that to happen.
Now for offering reasonable performance for new features, not only did both the NV10 and R100 offer features that performed at reasonable speeds, they saw
significantly faster adoption in the gaming community then any of the new DX9 features. Within a year of the NV10s launch almost all of the titles for best graphics relied heavily on features that were new to market in hardware a year earlier(Evolva, Sacrifice and Giants-HardT&L, Dot3 for a few off the big visual hits from 2K). Look at the situation now, we have that one port of a console title that is even close to comparable, and most people think CoD, MP2 and the like look a lot better. Pixel Shader 2.0 to date has been an abject failure in the marketplace compared to every new 'major' feature introduced in the last five years every way you look at it outside of marketing.
Sure, the 5800 was decent, and it was a good leap above the GF4. But the 9700 offered comparable performance in almost everything, along with nicer AA and the promise of far better DX9 performance. Most people don't upgrade cards as often as you, Rollo, so future performance may be an important consideration to some people. Granted, I've yet to see DX9 effects in games that blow their DX8 counterparts out of the water, but at least the possibility exists with 9700s.
The thing is it really doesn't exist. Check out the upper limits of what can be done with DX8 level shaders compared to what can be handled
in game by any current part at reasonable speeds. We have already seen Half-Life2's differences, noticeable but minor, and DooM3 using its DX7 era register combiner tricks still looks quite a bit better to me which is really the big issue. Most people do not understand what has been possible with hardware for a long time now, the problem was that the DX7 era cards were too slow to really show what the technology was possible of when pushed, same is true with DX8 era boards and the same is without a doubt the same about the current DX9 parts. They are simply far too slow, even the 9800XT, to come remotely close to pushing what PS2.0 is capable of.
But maybe I remain overly principled about DX9. Sadly, we still have to wait to see if DX9 speed concerns were justified. But 5800s and 9700Ps are still very fast cards, so it's not like their performance in new titles will be so slow as to be irrelevent.
Eighteen months. You take a board that launched with DX7 and move eighteen months forward in time you are looking at ~six months in to DX8's lifetime. Think about that. The R9700Pros shader performance was always a non factor in the real world, and that isn't looking to change within two years of its launch(if not longer). I'll tell you what Pete, I know you read a lot and you read a lot written by some very intelligent people. Some of the people you read thoughts from have a history when it comes to predicting the future 3D gaming market and they haven't been correct on a major feature shift
ever. You can check some of their history on it.
There was no chance DX9 was going to take off quickly, it was never a remote possiblity. This was painfully obvious to anyone who followed the gaming market instead of focusing all of their efforts on the hardware side. Simply look at what the market looked like at the time. Within a year of the launch of the DX7 era boards we had a selection of parts in the ~$100 street range that offered performance comparable, or superior, to the original 'high end' DX7 part, the GFSDR. Look at the launch of DX8 and move forward a year and we didn't have quite the same situation and ended up seeing slower adoption, although XBox ports helped mask this somewhat(still took longer then DX7). Look to the launch of the R300 and move a year forward and we didn't have a part close to $200 that could touch the R9700Pro's level of performance, and what's more we all knew that would be the case long before that point in time actually arrived.
From a developers standpoint what do you spend more time on, a feature that less then 2% of your potential market will be able to use, or one that will benefit the other 98%. Look at the feedback on Halo's visuals for evidence of what the majority of the market wants. Another element is that DX9 performance is far too slow even on the fastest part to see the theoretical limits remotely approached in terms of what the tech is capable of. Sure you can make a tech demo that impresses, Dawn blew people away(and that was on the 'bad' DX9 board), but getting acceptable levels of performance out of a DX9 class shader that is a big improvement over its DX9 counterpart in game with current boards? Very clearly the developers have spoken as here we are a year later and the only two titles we have to talk about are ports.
I'm running a R9800Pro right now and can say with absolute certainty that its DX8 performance is a much larger concern for me then its DX9 performance. It is far too slow running PS2.0, and too limited for that matter, to offer a substantial increase in visuals over what is possible with DX8 level shaders for real time gaming. Halo, TRAoD and now FarCry all back up the fact that the difference in visual quality is noticeable but slight, certainly not worth all the hype that some sites want to place on what is shaping up to be the most useless feature for end users to get close to the amount of hype it did.
Too many sites this gen have fallen in for the hype on PS 2.0. You take a look at recent reviews on numerous sites and you may seem something like 30% of the benches they are running PS 2.0 limited tests. I've seen Anand criticized numerous times because he doesn't replicate this incredibly misleading stance. You read a review and one board dominates 30% of the tests but they don't tell you that that actually translates in to
one game, and even then you won't actually want to play that one game anyway
ATi's PR department is likely loving it, and they have been utterly brilliant in their execution to exploit this(not an insult in any way, I honestly admire them here). They use a meak demeanor with, how can I say this in a diplomatic fashion, the sites that are very confident in their stances, and simply guide them to loudly boasting a direction they know full well the gaming market isn't going to take anytime close to soon(never correcting them of course, they would be fools to). For those that have a more realistic grasp on the gaming end of the market, they pull an event like Shader Day which has a dual purpose. First they got Newell hook line and sinker to go on the rampage about PS 2.0 performance which was a brilliant move on their part($6Million well spent), they used this in conjuction with his engine and nV's problems to put a perception in place that this game coming out real soon(within weeks according to what they were saying back then) would only run decently if you had a R3x0 based part in your rig. This had a huge impact on the community and even managed to convince a lot of the more down to earth sites that DX9 was going to be a big factor
soon. Brilliant execution on their part.
On the other side, nV went totally the wrong way with their architecture. If they were smart, they would have pushed the new lifelike lighting system their board was shooting for. By going this route, they could have set up a stand off of sorts between shaders and shadows and had a good rift in the community. Instead, they also decided to hitch their PR line on the shader stance. This was their biggest blunder this generation. Had they focused on their shadowing from Day one they could have then presented the DooM3 event as proof positive that they were right without having to say it for themselves(their marketing department needs a bit of work on subtle guidance). Because of this, they set themselves up for a comparison they couldn't win, and what's worse they did it for something that doesn't even matter(their PR department really fvcked up this gen).
Realisticly, the features that matter the most on the current boards is the ones we have been using all along. DooM3 and HL2 will hit along with FarCry- some sooner then others, but even when they do nothing we have seen indicates a big edge running the whiz bang DX9 shaders over their outdated DX8 counterparts. We can all tell the difference between running a game @1280x960 w/2xAA and no AF and running it @1600x1200 w/4xAA and 8/16x AF.
What this rant comes down to is this, the 5800Ultra, as an actual end product, had one
major problem, its loud fan. The rest of the perception of problems with the board surrounded sub optimal early drivers and the misplaced emphasis on DX9 performance. Is there really much of a difference between the 5800 and 9700 for end users today? Not really, certainly nothing resembling the perception of how the board was.