Originally posted by: Pete
The new memory architecture takes up
a lot of die space (it's the big thing in the middle, left of center).
A couple people at B3D have recently explained that the reason ATI does so poorly in D3 is because ATI's hier-z falls flat with Carmack's shadow algorithm, so ATI cards end up rendering a lot of unnecy pixels.
That sounds plausible because I heard Doom3 is designed with hypershadow in mind (ie a heavy stencil shadow load that NV40+ cards are designed to handle well). It seems that he didn't spend nearly as much time coding it for ATI's architecture; 'hypershadow or bust' if you will.
This may be a touchy subject for some (for those it is touchy for: get over it you fanboys
😉 ), but throughout the Doom3 development process, I basically heard that Carmack used Nvidia cards as a reference point, and hypershadow was a sort of give and take between him and NV - he planned on exploiting the feature heavily, and in return Nvidia emphasized it even more with Hypershadow 2 on the NV40 series.
This is in contrast to Valve, who wrote excellent DirectX 7, 8 and 9 paths for Half Life 2, allowing for multiple generations of cards to perform well in Source games. Despite their "ATI bias," Nvidia cards perform quite well on Source titles due to this more balanced approach, with the 7800GTX being basically the #1 GPU for Source these days, but good performance is all around, with no apparent chasm like with Doom3.
Other OGL titles, like Bioware's games, were probably (obviously, in some cases) written with nV in mind. So it may not all be ATI incompetance, but obviously early inattention by ATI (and 3dfx, etc.) was one factor that helped nV gain OGL preeminence. I suppose you could say ATI didn't pay attn to OGL b/c their employee make-up wasn't as much SGI or they didn't deem it as important, but Carmack was pretty close to God back in the early 3D days. I'd be surprised if they didn't spend some time on his games (which basically means OGL games in general). Nowadays, the proportion--or at least the popularity--of OGL is much lower, but you'd think Linux and the workstation market would compel them to put some ppl on it (which apparently they have, per the annual rumor, although the "rewrite" may not be as radical or as tangible as we may hope).
But it's not that much of a surprise to see an ATI card with a similar fillrate and hugely greater bandwidth compete with nV with AA+AF. I'd like to know what's up with Riddick, though. Given the manufacturing date of the review cores (late Sept and early Oct) and the apparently semi-trivial work on the MC that resulted in such huge gains, I guess ATI has some general tweaking to do with R520. The re-reviews comparing a 512MB X1800XT to a 512MB 7800GTX/Ultra should be even more interesting than I expected. 🙂
I just think that (as has been mentioned before), ATI doesn't put much of an emphasis on OpenGL at all, in their driver team, and in their product. Their bread and butter is D3D, and they are right to focus on that with the vast majority of games being D3D titles, but there's no question that there is some serious room for improvement in OpenGL performance, and they have made the issue even more sensitive by promising a completely rewritten OpenGL driver over the past couple of years, which has never materialized...
Regarding Riddick, I don't know exactly why ATI gets such abysmal performance, but I suspect their OpenGL drivers aren't helping the issue much and are mostly to blame. I could be wrong, though, as I'm not that well technically versed as many others (especially lots of the Beyond 3d regulars).
---------------
This ATI "fix" from Hexus sounds ok, but their results show nowhere near 35% improvements at any resolution; the best I can see is around 20% for a couple, and the results look sketchy. "Ultra quality" supposedly uses up to 512MB of RAM, meaning the NV cards may be unfairly penalized here since performance may be substantially better @ High Quality.
The results @ Hexus just don't look right - Nvidia usually has an even bigger lead in Riddick and Doom3 - those are ATI's achilles heel.
I'd like to see an Anandtech/Xbit comparison running HQ and Ultra Q with the final 5.11's before I make any judgements. I suspect that ATI's performance will improve, but I doubt this will be their "magic pill," and I predict Nvidia to retain dominance in Riddick and Doom3 at the end of the day.
Finally I have a question: is this fix supposed to be for the X1xxx series only, or does it apply to previous ATI cards as well?
