Anand Compares NV38/R360 IQ

Is AT going to put the cards in the lower tier into a third article?

Hopefully they will include some IQ comparisons for those cards as well, as it would help show if there are any optimizations being made solely to boost the top end's benchmark scores. The graphs in the latest article seemed kind of sparse without the cards that were featured in the first part.

 

According to the new 3DCenter.de article :

the new Deto 52.10/52.14 is worse than the Deto 51.75

when it comes to image quality; especially when using AF.


So I'm a little wondering how Anandtech came to his conclusions.

Link : http://www.3dcenter.org/artikel/detonator_52.14/


sorry for being so loud.

 

Is the pyramid demo in the 3dcenter article a custom demo? If so, there is a marked lack of improvement versus the standard timedemos.

Nice to see some screenshots that don't require me to squint.

 

Yes, it's a custom timedemo, i.e. a recorded MP match.

Sorry that there's only a German version. Should there be any troubles understanding the article (the images speak for themselves, mostly), feel free to ask questions here:

http://www.beyond3d.com/forum/viewtopic.php?p=177096#177096

Or open a new thread here (no registration required):

http://forum-3dcenter.de/vbulletin/forumdisplay.php?s=&forumid=77

93,
-Sascha.rb

 

Is the pyramid demo that much more demanding on the system than the antalus flyby?

The performance figures seem pretty steady between the scores of the pyramid levels, but the antalus flyby has a dip in the no AA/AF portion of the 52.14 dets. I wonder what causes that dip.

In a side note: The full size screenshot links are working in the AT article. It's just too bad only a select few programs have that option. What's up with the lack of Tron screenshots? Perhaps IQ isn't a concern according to them, but I like screenies.

 

Much, much more demanding. It's like a botmatch run with no botmatch AI, but with all the models, shadows, explosions, gunfire lighting etc. Have a go at it yourself:

http://www.3dcenter.de/downloads/ut2003-pyramid2003.php

There's still some CPU usage due to the physics engine, but basically, the Pyramid demo is mainly GPU bound.

93,
-Sascha.rb

 

Look at this objectively - ATI won 90% of the benchmarks in there, and some by a not inconsiderable margin, and 90% of the conclusion was apologising about NVIDIA's performance, even with the suggestion that they may be faster at some point in the future. I'm sure the single line "ATI is still our recommendation" is of so much joy to the ATI hardware and software engineers who've spent their time making hardware that beats NVIDIA in 90% of the games tested (and without IQ degration).

 

Quite lengthy, but I was rather hoping for a broader range of Dets and Cats to see play and perhaps some refrast examples. And a LOT more full-size image testing.

Sadly the whole matter is MUCH more than a pair of reviewers can do, so we're going to have to wait on 52.14 to run the gaming gamut and pull specifics from ALL the sites out there. Thankfully there are a lot of amazingly anal observers out there, so given a few weeks and a LOT of reading we'll have the best idea.

 

That is too funny... If that isn't a sig-worthy quote, I don't know what is!

 

Eh ?? Well that is simple..3Dcenter know what their talking about

 

OK

Without regard to gains being relative to achieving the same output, I do think this is a fair objective assessment.
But it seems to me that if this is done without achieving the same level of output, such "performance gains" is best characterized as "fictional". Therefore, it seems important to me to take a reasonable amount of steps to answer that question. I think AT failed.

Hmm...I don't quite get how you propose this statement as factual, such that you make it a fundamental premise of an "objective" evaluation.

I have a problem with this statement being proposed as such due to an abundance(!) of factors:

• A concentration on the Anti-"Synthetic Benchmark" mantra, while completely ignoring the "fixed camera path problem" that is the real issue nowadays, and which is more likely to be exposed with dedicated ("Synthetic", nevermind how representative the workload) benchmarks. I feel like someone is trying to sell me a bridge, and some reviewers seem to have bought it already.
• Selective presence of full screen images for reader controlled comparison. It takes work to reduce images after capture, so it isn't like they didn't have them to begin with. It also seems to me that you can "cut and paste" any "click to view original" code, so it wouldn't be more work to include them. This leaves me with the impression of selective truth to facilitate Anand's conclusion in place of the reader being given full information to come to their own, which in my experience is one effective tool of "marketing" and/or "lying". The rollover issue someone else mentioned also concerns me, as well as the idea of jpeg based comparison of filtering and AA in still screen shots...especially with so many being only available as reduced to 450x450 (as far as I could determine).
  [list:49e5e70d81]
• Aquamark 3 : NO full size image.
• C&C Generals : full size images.
• F1 Challenge : NO full size images (with an AA/AF comparison, even!).
• Gunmetal : NO full size images.
• Halo: full size images (NO AA/AF comparison...no evaluation, illustration, mention, or consideration of shader effects at all, which you think might be warranted with the large amount of confidence Anand speaks invests in the 65% performance gains. Ah, but even though they can relate to graphics workload, cut scenes don't represent gameplay anyways...but do the screenshots? What about representing compiler technology or card performance, as Anand uses them for (especially in his conclusions about compilere technology)...which relates to the fixed camera path issue I'll cover in more detail later.
• Homeworld 2 : NO full size images (with an AF comparison...AA didn't work for the nVidia cards, though resampling to 450x450 seems likely to make that a bit irrelevant for the screenshots).
• JK- Jedi Academy : full size images (and finally with an AA/AF comparison. But hold on, they recommend you set them in game, but then went and set them in drivers, though in game settings are exactly where a difference should be most evident between ATI and nVidia...if it weren't a jpeg still screen shot and there were more than one texture layer used in the game (?). It is the only full size image AA/AF comparison that it seems possible for the reader to evaluate effectively, yet their selection of how to test it doesn't seem logical or consistent with even their own statements, nor with issues surrounding nVidia and ATI in general.
• NWN : NO full size images (with another AA/AF comparison).
• SimCity 4 : NO full size images (with another AA/AF comparison).
• TR-AOD: The most competent and thorough explanation in the review, AAICS, and full size screenshots offered for examining PS 2.0 output. What disturbs me is the lack of an actual thorough PS 2.0 performance comparison between ATI and nVidia (from any benchmark, it wouldn't have to have been TR-AOD) to go with what they use TR-AOD to show. Least logical comment in the discussion : "I know everyone will want to take these numbers and say that they universally describe the DX9 performance hit on NVIDIA hardware, but we have had plenty of other benchmarks today that show very different results." In contradiction to his discussion of how specific to TR-AOD his analysis is, it is even used to validate every other shader benchmark in a way inconsistent with the data shown.
• UT2k3: NO full size images (and Yet Another AA/AF comparison at the reduced size). No controversy around this game and filtering...let's just show the full size images for JK instead. Did I just miss the full size images being offered for this? That would make more sense than what I think is the case now.
• WC3: NO full size images (with an AA/AF comparison).
• Wolfenstein-ET: NO full size images (with an AA/AF comparison).
• X2 : Not sure what the screenshots really mean. Looks like something in the balancing act I've thought AT has tried before, where (IMO) the "alternate" IHV is shown in the best possible light in one targeted instance to shield from accusations of bias, even when that requires doing something that doesn't even seem to make sense. ATI's AA is better, but what does showing us a screenshot of a high contrast logo demonstrating this really tell us, when you propose it represents a game? Why not a screenshot from the models/graphics for the rolling demo?

OK...maybe I'm looking at old pages, or a complete collection of full sized images were offered elsewhere and I missed it. Perhaps someone could correct me if such is the case, or I made some errors in the above. As it stands, I'm just a bit boggle by the inconsistencies, and that still leaves the following...
[*]This next list is made especially poignant In the context of the above, and conclusions like this: "The very large performance gains we saw in Halo, Aquamark3, X2 and Tomb Raider can be attributed to the enhancements of NVIDIAs compiler technology in the 52.xx series of drivers", when this relationship is made up on the spot in the conclusion.
A lack of answer to the some of the specific questions (borrowed from Valve's presentation, and related outside of that context as appropriate) seems a gross oversight:

• Camera path-specific occlusion culling.
  Independtly confirmed to have been done by nVidia before. On the camera path, image quality is (usually) identical, but completely unrepresentative for performance evaluation or comparison. Big problem when this image quality comparison doesn't even seem to take the rudimentary step of custom timedemos to begin to try and avoid this happening.
• Visual quality tradeoffs (lowered filtering quality, disabling fog).
  Hmm...I'm pretty sure I've seen "4x.xx" drivers without filtering issues, though the fog question is still unanswered for me. I think Anand could have answered it, but I don't recall him trying. This could, however, be something that the new drivers improved.
• Screen-grab specific image rendering.
  This remains unconfirmed outside of Valve's statements, AFAIK. Does seem to throw another monkey wrench in limiting the selection of AA/AF comparisons that users can effectively evaluate for themselves, while at the same time proposing they are universal indicators of "improved image quality" and skipping consideration of other AA/AF issues in your conclusion. As if there weren't enough wrenches in that already.
• Lower rendering precision.
  Big problem, where was the investigation of this? Perhaps they meant B3D to be used for this, but the comparison they linked to doesn't seem to answer that question. Nor do they seem to consider it for their conclusions.
• Algorithmic detection and replacement.
  Actually, with the HLSL compiler SINCOS issue, I think there is a valid case for nVidia to do this at the moment. Unfortunately, I don't think nVidia would limit themselves to validly addressing such problem when they did do shader replacements. Of course, if you use Cg and don't get performance gain compared to HLSL, this should remove that concern, AFAIK. In any case, see the above issue with where I think AT went wrong with evaluating the concerns this brings into it.
• Scene-specific handling of z writes.
  I'd say this is a big problem in reality, but not a big problem for AnandTech failing to check this (at the moment, with the lack of tools we have, or at least that I know about).
• Benchmark-specific drivers that never ship.
  I find certain displayed "logical conduct" regarding driver usage policy at AT some reason for major concern. However, this is only a big problem if the drivers used by Anand are actually never released, despite "assurances". This could end up being moot, however, if they do so...except, of course, for the "HL 2 performance update" blog entry (which, however, wasn't actually in an article, though the conclusion to this one reminds me of the same mindset).
• App-specific and version specific optimizations that are very fragile.
  I think this is more of a "big problem" for a game that is going to be updated very frequently, when performance will vary greatly as these optimizations "break", unless the developer agrees to freeze their shaders for the convenience of an IHV making custom replacements . I know this is/was a big problem for HL2, but I don't know how much of an issue it is for other games...except as far as having to wait for nVidia to get around to your game, and making patches a possible nightmare for developers. Also, the uncertainty as to whether nVidia will be able to have success in creating an equivalent shader (outside of those games and timedemos used as popular benchmarks) or not.

[/list:u:49e5e70d81]

I don't really think much of this article, overall...quality over quantity. There seems to have been a good opportunity and frame work for some really good work here, but with spotty follow-through, though a lot of time seems to have been spent. The use of reduced size images seems ridiculous to me (I'm still wondering if I just missed something?), and even if that wasn't an issue, there is a complete failure to addresse even the simplest of the last list of concerns, as well as a significant lack of coherency in the "logic" present in the conclusion.
Where there is full sized IQ analysis, the choices of methodology and performance comparison presentation seem to be reasonably characterized as being chosen selectively to prevent showing nVidia in a bad light, and then applied as a universal basis for evaluating all other information, even where the same things were specifically not evaluated in what it speaks about (AA/AF and the relative performance of the "PS 2.0 compiler" between nVidia and ATI).
The (absent) performance comparison (not the same thing as "performance hit relative to itself comparison") between ATI and nVidia in TR-AOD would seem an excellent opportunity to test the idea of nVidia's ability to "take the crown" from ATI as they propose in their conclusion, yet its not even clear to me if this comparison is without shaders completely, or with minimum settings. Not that I could tell anything from the screenshots anyways (way too dark), but I'm assuming it is due to my monitor that I'm replacing.

This, to me, is large scale deception and/or significant incompetence. Perversely, the strength of what they did show with TR-AOD (I don't have an issue with it, I have an issue with what it is used in place of and the information that ends up being witheld due to that) seems to argue against the one of those I personally find the least disappointing.

Where do you get the idea that nVidia is "hated" because they are behind? Maybe some diehard nVidia fans "hate" them for that (?), but I think the hostility towards nVidia I'm most familiar with is for reasons besides that.

The issue isn't as simple as "the reported fps gains are greater" and "Anand says the image quality is improved".

That conclusion does not seem even remotely objective to me, nor does it seem to have been determined based on objective evaluation.

To be fair:

I know there are some people who go to extremes in hostility, but they happen to have been given a rather large amount of cause. For instance: I didn't generally find it "scary" that people condemned ATI vehemently for "Quack", and there was quite a bit less behind the issues going on then. I did, and do, find some of the "logic" surrounding both issues a bit "deficient", "lacking", and therefore "surprising" to me in that way...but that has to do with (some of) the reasoning, not the hostility itself.

My concerns above relate to both matters.

 

First of all, the small images are not resized/resampled, only cropped. You can see aliasing artifacts on them well enough... assuming that there are any.

Also, there aren't any surfaces that would show any serious texture aliasing, because of the camera angle. Useless for AF comparison.

However, the lack of contrast and detailed patterns in the floor/wall textures basicaly hides any texture aliasing / bluring. Useless for AF comparison.

See Homeworld 2.

Once again, there are practically no surfaces in the view that would show any texture aliasing. You can still see that something isn't right with the NV5x drivers, but there's not enough information to make any observations.

See Homeworld2. The camera angle wouldn't show any aliasing... also note that they've originally mixed up the NV5x images, as I've previously noted. And if you actually play the game, the texture aliasing is very disturbing, especially in cutscenes.

See UT2003.


Conclusion: the majority of the screenshots are simply useless for any image quality comparison. Some of the OK images show noticeable problems with the NV5x drivers.
The more I look at this article, the more disappointing it becomes. I seriously hope that B3D and ExtremeTech will get their hands on the new drivers + a GFFX ASAP.

 

Phew. I'm glad this was me missing this...resampling would have been a bit "remarkable". Picking portions of the screen while deciding to selectively refuse to offer the original screen images is quite problematic enough. I guess squinting would have paid off after all, at least as far as evaluating what was actually shown, and outside of any resolution disparities.

 

ROFLMFAO~~~~

Best line I've read this morning, thanks for the laugh...could you please explain that to Evan Lieb next time you see him? I'm having the damnest time trying to beat some sense into that fellow.

 

A nice guy at AT forum, hominid skull, has put an animated Gif of the F1 images. http://www.aclift.eclipse.co.uk/example2.gif

I'm suprised AT didn't do this as it's alot easier to view any differences - even with those tiny images.
AT's conclusion should have included that if you don't want line crawling even with AA on then of the reviewed cards ATI is your ONLY option right now.

 

Wow, I thought it was just AA that was ATI's big strength in F1, but looking at the white road line, you can see that NVidia compromised AF as well. I guess this confirms the results of the 3Dcentre article. Doesn't it look like NVidia is only using 2x horizontal AA? If it was 4x (even ordered grid), shouldn't you at least be able to see one medium blue color on horizontal surfaces, like the Michelin sign?

Anyway, I always found that racing games are where you need AF the most, and it can even improve your racing ability if the road lines disappear without it.

 

I don't think Anand's goal was really to show the differences, I think he was trying to hide them...thus the tiny screenshots.

 

Pyramid2003 benchmarks with 8xAF + 4xAA:
http://www.3dcenter.org/artikel/r9500+9700+9800_vs_gffx5200+5600+5800/zj_bench_ut2003_d.php

Without AA and AF the Pyramid2003 timedemo ist CPU limited on highend graphic cards. But with AA and AF it is highly gfx limited (7,5 fps on GFF5200). Anyway, this demo wasnt taken to show a CPU or a gfx limitation, it was taken to show UT2003 real gameplay performance.

UT2003 Timedemo with more gfx limitation:
http://www.3dcenter.org/downloads/ut2003-overkill.php

 

Was it just me, or did a couple of the side objects look a little sharper under the old dets over the cats? Couldn't say, since I'm not sure what they're supposed to look like

 

the dets show much more blurriness than the cat 3.7 if you look into the distance.. there are also very apparent 'boundaries' where filtering is being applied it seems..