State of the Graphics Industry Rant

[darkblu]

freakin' munchkins, Russ! when my t. pratchett reserve gets all depleated i could get my fix just from your posts here

 

[Joe DeFuria]

Regardless of whether its fixed now, it was definately less capable then when compared to the competition. It didn't filter at 45 degrees off the UV axis.

It was also much FASTER when compared to the competition.

No one is saying (at least, I don't think so), that just because something is different, that makes it OK. However, when something is different, you must look at both the benefits and the drawbacks.

You keep on saying 8500's implementation was "less capable." I could say the same for the GeForce 3/4's. The difference being, GeForce was "less capable" in the performance department.

Or should they say, "each card is special in its own way, and momma loves them all the same"?

As I said....they should say "each card is special in this way, and not-so-special in this way." And go ahead and throw in personal opinion, like "Even though the 8500's method doesn't technically give as good a quality result, I don't notice the difference in game play. So the performance benefit is worth it, IMO." or "Even though the 8500 is faster, it's not that much faster that it makes up for the shortfall in image quality, IMO."

To me, they're there to compare as many different products as they can FOR ME and give me their feedback, both with objectivivity and with subjectivity so I personally don't have to see them all with my own eyes.

Agreed. And the "objective" part cannot be achieved if the reviewer only looks at ONE aspect of a featue in a vacuum (image quality), without considering other aspects, like compatibility & performance.

 

[jb]

Regardless of whether its fixed now, it was definately less capable then when compared to the competition. It didn't filter at 45 degrees off the UV axis.

While that is certianly an issue I saw no one mention it unitl I provided that SS:SE screen shot. Then it showed up on forums and other places

However the reason why I think people thought it was flawed steems back from an old (FS maybe?) interveiw where ATI said it used an adaptive method and thus has a smaller pef hit. Once that happened I saw a lot more people in the forums say ATI was cheating, ATI is not giving us real AF, ATI AF is flawed, ect... Also the TR and other places showed that its not doing trilinear when its doing AF. Once again the noise started in the forums, ATI is cheating with AF, not doing full AF, ect.....

I also have not seen an reviews, up until MiceC made a one line comment in his GF4 reivew, about MSAA and alpha textures. So yea some reviewers are TFC

 

[RussSchultz]

(off topic--as a data point, is the 8500 still faster, now that its output is 'fixed'?)

(on topic)
I'll have to say, I can see several sides:
1) The reviewer should at least try to present all the facts. (I.e. yes, display is worse, but speed is better)
2) On the other hand, when do you draw the line and say "hey, this method is cheating"?

This is a very fine line that is funny as hell to watch reviewers take one side of, and then have fanatic-boys come back and blast them for being biased. And then flip-flop over a similar but reversed situation--all based on who's ass you think is a better smooch.

(p.s. Joe, I don't think we're on fundamentally different wavelengths here.)

 

[Joe DeFuria]

2) On the other hand, when do you draw the line and say "hey, this method is cheating"?

I'm not so sure that "cheating" should ever really be used to describe a method or implementation. IMO, using that word should be more reserved for PR statements, or reviewers that make unfair apples-to-oranges comparisons.

For example, is nVidia cheating by having Quincunx AA? I don't think so. However, I would consider it "cheating" (or dishonest) if someone posted Quincunx benchmarks against someone else's "standard" 4X AA implementation in an effort to prove which is "faster", without commenting on image quality.

 

[Doomtrooper]

It is silly not to compare features because they do it differently, as for anisotropic filtering there is no Standard so to speak, Texture compression there was. Looking at written documentation about Anisotropic filtering shows it was doing what it was supposed to do on a 8500.
Supersampling has been used against MSAA and MSAA has been used vs. FAA Fragment Aliasing...to say we are not using a Radeon 8500 because it does AS filtering differently is just plain STUPID as none of the cards ever implement these features the same.

 

[darkblu]

It is silly not to compare features because they do it differently, as for anisotropic filtering there is no Standard so to speak, Texture compression there was. Looking at written documentation about Anisotropic filtering shows it was doing what it was supposed to do on a 8500.

wait a sec! 8500's aniso was flowed even in the context of its own unique implementation. that's why it got fixed in the following versions. please, let's call things their real names. helps for keeping one's sanity.

 

[RussSchultz]

Would only implementing anisotropic filtering on the center 1/3 of screen be a cheat? (not that I know of any card that does this)

I would think so, even though the output probably wouldn't suffer much on most games out today.

What about a card that would render in 7:7:7 when 24 bit mode was chosen? Or 6:6:6? 5:6:5?

Or a card that does 9:9:9:5 instead of 10:10:10:2? The user probably won't notice the difference in normal circumstances, but it isn't what is advertised.

Along those lines, I think that if the 8500 did have issues with not applying the filtering properly when the texture was oriented in a particular direction and did this to gain speed, that also would be an over-optimization. (Otherwise, it was just broken)

There's obviously a line somewhere in that gradient of optimization that eventually turns to over-optimization. I'm not going to put a finger on it, but to quote a famous senator: I know it when I see it.

But, just to poke at Doom a bit:
The 8500 AF did what it was supposed to(except in some cases). Its ok to say that other part quietly.
The GF series of cards did meet the exact specification for DXTC(even though its output was out performed by other cards). Same here.

 

[Doomtrooper]

You are also using reviewers opinions who chose to use the 'word fixed' vs. improved upon...

ATI improved their adaptive algorythm to adress these minor issues, the same can't be said about texture compression as it still isn't Fixed.
In 99% of the General use the Radeon 8500's implemention did not effect image quality, actually improved it...

 

[Doomtrooper]

But, just to poke at Doom a bit:
The 8500 AF did what it was supposed to(except in some cases). Its ok to say that other part quietly.
The GF series of cards did meet the exact specification for DXTC(even though its output was out performed by other cards). Same here.

Lol its funny watch the way people think on this subject...

1)The Geforce cards didn't meet spec at all on texture compression otherwise we wouldn't need third party tweaks to hide the flaws would we..
2)Show me the documentation (the standard) that ATI didn't follow with their implementation ?? Its not that hard to prove

 

[Joe DeFuria]

Would only implementing anisotropic filtering on the center 1/3 of screen be a cheat?

IMO, no. It would be prudent for any reviewer to comment on image quality in any comparison though.

What about a card that would render in 7:7:7 when 24 bit mode was chosen? Or 6:6:6? 5:6:5?

Well, is it considering cheating today or in the past when "32 bit" 2D cards rendered in 24 bit?

EDIT: I know that's not completely analogous (32 bit was usually referred to as "true color mode". But some PR depts "cheated" when they claimed their "32 bit" was superior to someone else's "24 bit" in terms of quality.

If 24 bit mode is chosen, then 24 bits should be used. 24 bit has a clear, unambiguous meaning.

Or a card that does 9:9:9:5 instead of 10:10:10:2? The user probably won't notice the difference in normal circumstances, but it isn't what is advertised.

Ahhh...that's the key. "Advertised." Is the card advertised to support 10:10:10:2? Then yes...actually doing only 9:9:9:5 is "cheating" or more accurately a lie.

I think that if the 8500 did have issues with not applying the filtering properly when the texture was oriented in a particular direction and did this to gain speed, that also would be an over-optimization.

You can't reasonably come to that opinion without knowing one crucial factor: cost. Assuming it was a chosen optimization (and not a hardware bug), you have to consider how much it would have "cost" to apply to filtering as the texture was rotated.

In short, you are looking at this purely from performance / quality perspective, whereas the designers must factor cost into the equation.

There's obviously a line somewhere in that gradient of optimization that eventually turns to over-optimization.

Agreed. The thing is, that line is going to vary not only from person to person, but from game to game. R 8500's aniso may be perfectly acceptable to the majority of FPSers, and unacceptable to Flight Simmers at the same time.

That is why I would prefer these things to not be labeled as "cheats". Simply explore the consequences (good and bad) of the implementation. "Cheating" has some global, all-encompassing meaning. Nobody says "this cheats with flight sims, but doesn't with FPS games". It's cheating or it's not.

 

[RussSchultz]

The S3TC specification has absolutely nothing that defines the manner in which interpolation of the generated colors must be done. To do it in 16 bit space is just as 'valid' an interpretation as to do it in 24 bit space. One obviously has better output, however.

And as for the anisotropic filtering on the 8500, you're right--if thats their implementation, you can't say they did it incorrectly (since by the circular argument, their implementation is correct according to their implementation). At that point, I'll have to say their implementation sucked, since it failed to work when the texture was oriented in particular ways.

But on other topics, it looks like we might get some rain here in Austin today. I'm hoping we don't since the house isn't completely weatherproof yet.

 

[Doomtrooper]

The tech-report posted these findings on the R200's launch...

Figure 14: R200 anisotropic filter, highest quality (left) and high quality (middle)
GeForce3 anisotropic filter (right)

Anisotropic filtering reduces the too-fuzzy or too-sharp filtering that occurs with isotropic mipmapping when a pixel maps to a rectangular region in the texture space. The anisotropic filter employed by R200, in terms of texture clarity of distant objects, is conclusively superior to GeForce3 (Figure 14). A slight disadvantage of R200 is that it does not filter between mipmap levels (i.e. trilinear filter), though this is not noticeable with a high level of detail.

Man, and ATI's implementation is inferior

 

[RussSchultz]

I dunno, Joe. I prefer correctness of result, personally, though I realise the entire industry of 3d graphics is built around optimizations of one sort or another.

It certainly makes it hard to compare one product to another (unless you've already got your mind made up) when there's so many varied "optimizations" that swing back and forth.

I'd prefer a case where the default settings would be bit exact output between cards, and let the user choose the optimizations he wants or not. Of course, that's a complete pipe dream.

 

[Joe DeFuria]

To do it in 16 bit space is just as 'valid' an interpretation as to do it in 24 bit space

Agreed. "At that point, I'll have to say their implementation sucked."

 

[RussSchultz]

[paraphrase="Doomtrooper"]My card is better than your card
[/paraphrase]

Could you please take these endless comparison discussions elsewhere. This conversation really has nothing to do with NVIDIA vs. ATI, other than they are convenient examples.

 

[RussSchultz]

To do it in 16 bit space is just as 'valid' an interpretation as to do it in 24 bit space

Agreed. "At that point, I'll have to say their implementation sucked."

Thank you, Joe, for leaving off the next sentence which essentially said exactly what you did.

The wonders of selective quoting never cease.

 

[demalion]

Would only implementing anisotropic filtering on the center 1/3 of screen be a cheat? (not that I know of any card that does this)

I would think so, even though the output probably wouldn't suffer much on most games out today.

What about a card that would render in 7:7:7 when 24 bit mode was chosen? Or 6:6:6? 5:6:5?

Or a card that does 9:9:9:5 instead of 10:10:10:2? The user probably won't notice the difference in normal circumstances, but it isn't what is advertised.

Along those lines, I think that if the 8500 did have issues with not applying the filtering properly when the texture was oriented in a particular direction and did this to gain speed, that also would be an over-optimization. (Otherwise, it was just broken)

There's obviously a line somewhere in that gradient of optimization that eventually turns to over-optimization. I'm not going to put a finger on it, but to quote a famous senator: I know it when I see it.

But, just to poke at Doom a bit:
The 8500 AF did what it was supposed to(except in some cases). Its ok to say that other part quietly.
The GF series of cards did meet the exact specification for DXTC(even though its output was out performed by other cards). Same here.

For my own take (aka opinion) I'll say:

1) I agree with Doom to a degree in that the DXTC problem is a different issue. I even agree with his evaluation about it being much more of something "broken" than 8500 aniso. I had thought that the drivers always mapped that mode to one that looked better but was slower. In that case, it is OK not to mention it at all in a review unless design flaws are a topic of discussion, as long as the benchmark and IQ comparison are not skewed to circumvent the issue this provides (i.e., benchmarks with texture compression off and then comparing IQ and performance).

2) We should be comparing GF anti-aliasing feature set to 8500's aniso feature set. Both have shortcomings in specific situations, and both offer advantages. Going from there and labelling the tradeoffs universally "broken" is subjective...it can only be stated they are "broken" in specific objectively defined cases, and there seems to be some confusion that these two uses of the same word are synonymous.

The way the 9700 fits into this seems to be confusing the issue. It DOES make the 8500's aniso look broken in comparison by eliminating its tradeoffs, but it also makes the GF 3 and 4 aniso look broken in comparison for the same reason. This is separate from the 8500 aniso and GF 3 and 4 anti-aliasing comparisons, however.

 

[Doomtrooper]

[paraphrase="Doomtrooper"]My card is better than your card
[/paraphrase]

Could you please take these endless comparison discussions elsewhere. This conversation really has nothing to do with NVIDIA vs. ATI, other than they are convenient examples.

WTF are you talking about ?? That is what this thread is about, one item was why wasn't the 8500's anisotropic filtering used in tests when evidence like the above proved it was superior...

Typical for Russ...start getting proven wrong and resort to personal attacks..I have some news for you Russ.'You don't know everything

 

[Joe DeFuria]

I dunno, Joe. I prefer correctness of result...

Wouldn't we all! And we'd also like the results at infinite speed, and sold for $10 MSRP. That's the trouble with consumers, and that's what product managers and engineers face every day.

It certainly makes it hard to compare one product to another.

Sure does. That's why the majority of "comparison" reviews are complete crap. Because it takes time and some thought to do it properly.

I'd prefer a case where the default settings would be bit exact output between cards, and let the user choose the optimizations he wants or not. Of course, that's a complete pipe dream.

In practical terms, it's more like "a waste of money." How about "default AA"? Must every card support ordered grid super sampling just to have some base line comparison even if that mode will never be used? The people who buy these things are not interested in them for the "technical merit". They are interested in getting "an experience."

Since "the best" experience of zero cost, infinite performance and "technically correct" quality is not an obtainable goal, trade-offs have to be made. Someone's going to decide that "this feature" should be dedicated more resources to it, because the target audience cares about it more, and someone else is going to decide that this other feature will get minimal resources, because consumers "may not even notice" if they spend more to do it "right."

The BEST actual products, are the ones that choose the "correct" trade-offs for you. Keep in mind that the best trade-offs for your, aren't the best for someone else....