NV preps 9800GTX+ against HD4850

It never ceases to amaze me the amount of flip-flopping some people will go through to catch up with their favourite IHV's PR.

32 bit precision is great then it's not. FP16 is great then it isn't. Latest shader model is great then it sucks. High level of AA is excellent then it's not needed.

If anyone feels hurt by that, please note I'm talking about both camps here.

 

I like using 16xCSAA when the performance cost is minimal. If the drop from 4xAA to 8xAA on ATI hardware is as minimal, then I'd likely use that mode.

I'm somewhat disappointed and excited about the current round of products. I'm somewhat disappointed at the current cost of the GTX 280 and as of yet lack of new features. I do like the performance increase without having to deal with possible SLI issues. I'm somewhat excited at the performance levels of the various AA modes of the RV770 and the possibility of the HD 4870 X2 not being typical AFR solution and likely cheaper than a GTX 280. I will be disappointed if it's another typical AFR solution.

I like the exceptional values being presented to the consumer, a HD4850 at $170 - $200 level and a 9800GTX/+ at $200 - $230 levels.

 

I don't see any parade of people yet. Certainly I attempted to correct what I believe to be a serious misconseption in the original post - i.e. that the quality of 16x CSAA is unequivocally better than that of 8x MSAA. I think that that is a statement that is hard to defend either objectively or subjectively.

I have no problem with 16xCSAA as an AA mode, and in none of my comments have I tried to brand it as in some way worthless or unworthy of consideration. There does appear to be one person here, namely you, making decidedly negative statements about 8xMSAA -

"Advantages are minor and abstract"

I think that that is a statement that is just as difficult to justify (certainly to someone conversant in the subject matter) as the statement that 16xCSAA is better than 8xMSAA.

"Subjective and questionable" as well... I guess that's similar to "minor and abstract" again, yes?

I'm really not sure why it's necessary to react so negatively to 8xMSAA (which offers clear and consistent improvements over 4xMSAA). I dare say that 8xMSAA or 16xQ CSAA are also wonderful for nVidia users in the applications in which the performance is good. It's always nice to pick the highest level of AA that gives the performance that you desire.

I tend to agree with the basis of the sentiment, although I certainly don't think the sample density of 4xAA is sufficient. We would probably like to go quite a bit higher than an 8-sample density ideally, but memory can start to get a bit prohibitive.

I don't think that 8 full samples is at all unreasonable given the amount of memory on current cards.

Actually someone was talking about that - me. That's how conversations and discussions start - by someone mentioning something. I'm not sure why a simple statement about AA with custom filters as another alternative method requires a defensive response...?

Anyway - since in your statement above you come down firmly in favour of increasing quality without using much more memory you should like the custom edge-detect filtering on AMD hardware a lot, since that's exactly the result that it achieves.

At what level does AA become silly in your estimation? At 4x, 8x or 16x, or somewhere else?

Yes, from a pure quality standpoint, and ignoring performance I would like to have supersampling as well. Of course, at reasonable levels of sampling it start to get extremely expensive.

Not just "subjective" "questionable" "minor" and "abstract" then, but 8x MSAA is actually "worthless"...?

That's a really strong statement. I certainly wouldn't feel comfortable about backing up such a statement if I was talking to other graphics professionals about anti-aliasing, but maybe that's just me.

 

You seem to be under some false pretense that I am reacting in any way. I have been using 16x CSAA since day 1. Turning on 16xQ has had its fair share of problems. ((8xQ included)) with transparency AA and memory footprint. with transparency SS and alot of alpha's on screen. 8x TSS becomes more of a performance hog than 4x SS. Which makes me question the usability of such a feature. Adding into that that at high resolutions. You really can only get away with the 4xAA memory footprint to begin with. I could say your defensive about Nvidia users likely preferring CSAA. But come on. Lets be adults here. I have good reasons for thinking 16x CSAA is an incredibly valueable mode and that 8x multisampling in comparison has too many problems to make it worth while on Nvidia hardware.

I think your nitpicking for the sake of nitpicking here. While I do agree that increasing AA quality at the cost of memory is important. But so is overall performance. If I lost a significant amount of performance using 16x CSAA over 4xAA. I wouldnt use it either. Remember your not trying to convince me here. Your trying to convince users as a whole. That these modes add alot of value. If they cost too much performance or run out of memory. Then the modes lose value.

I have 5 GT 200 type variants of cards. With at least 896 megs of memory. I have a bit more memory to waiste on 16xQ/8xQ. Even with this in mind. I dont use the modes because the improvements they provide ((if any because they dont always add quality)) significantly effect z-fill rate performance and transparency SS performance.

Most users with 512 cards. Since we're talking about 512 meg card thread. Dont really have that luxury. Even at 1600x1200 stuttering can be observed with 8xQ AA on many modern titles due to the memory hitching. And as the resolution increases this becomes even more of a problem. Adding to that. Video Memory usage in titles is going up. Not down.

Maybe if you dont have the fillrate. I could easily argue that 16xS/8xSQ/4xS/8xS are not terribly expensive in 95% of the games out there. But that doesnt make them terribly worth while. From a quality perspective the gain is minimal other than the improvement to overall scene quality. Texture, Shader, And Alphas. 1x2 SS and 2x2 SS will provide noticable Quality improvements to the shader aliasing and LoD adjustments that come with Nvidia's supersampling.

Please. I'm not getting into this subjective argument with you. Why dont you go and tell everybody that they need 8xAA on the forums. Or heck start a marketing campaign about it. See how people react. You won't convince me. I've measured, tested, examined IQ on this for ages now. My subjective views on what I deem as usable arent that important.

Whats important is that what others deem usable.

Since I was discussing transparency SS specifically. Yes worthless. Any scene with loads of alpha's using 8x multisampling and transparency SS is dog slow. Slower than just doing traditional Supersampling. When the cons outweigh the gains then I consider it pretty worthless.

Take serious sam 2 for example. Try turning on 8x multisampling and transparency SS and your performance will drop to dog levels because its an entirely inefficient way of providing AA too alphas. If the performance hit from transparency SS was smaller. Then 8x would be slightly more valuable in my eyes. Hence the need to allow for controllable levels of transparency AA. And I'm not just talking about halfing the SS and getting extremely unideal sampling positions. I'm talking about True 2x/4x/8x transparency AA seperate from the base multisample value.


With 16x CSAA. I at least know that I am not forcing a huge fillrate bottleneck on myself because only 4x transparency AA is called on with transparency SS. 8x TSS is simply too slow in titles that use alot of alphas and IMO brings questionable benefits.

Chris

 

Not true.
Lots of stores everywhere list HD 4870 in 25-28 June. I´am sure that on those dates, the local stores will have the cards.

The HD 4850 started to sell on 17 (instead of 25) and HD 4870 will start to sell on 25 instead 8 July.

There are plently HD 4850 in the market and the launch date was suposed to be on 25. Same will happen to HD 4870.
Probably the production of the cards are beating expectations, so they started to sell much earlier.

 

Fair enough - I apologise if I was proceeding under the false assumption that you were reacting defensively.

Possible, I suppose, although I don't think I would view this as the norm by any means.

Based on my experience, with decent memory management on the part of the driver, you can get away with an 8xAA memory footprint at resolutions up to (and often including) 2560x1600 in the majority of games. Certainly 1080P or 2048x1536 rarely present a problem.

Don't get me wrong though - there are certainly exceptions to this (Crysis for example)

You could, but I think you would have little grounds for doing so based on my responses in this thread - at no time have I implied or suggested that nVidia users would be better served by enabling 8x MSAA on their hardware if they take a large performance hit in doing so. Nor have I attempted to present the characteristics of 16x CSAA in any unwarranted bad light.

And I don't deny that that may be a reasonable position with respect to nVidia hardware. It's if you try to imply that it's a superior solution in general when compared to an equivalently quick 8x MSAA implementation that we will end up disagreeing.

I'm not trying to convince users as a whole - where did you get that idea? I am posting technical information, about a technical subject, on a technical website.

My original post was to correct what I viewed as a misrepresentation - that's all. I know that I've now ended up debating various pros and cons with you, nothing wrong with that, but I'm hardly on a crusade rampaging around the internet trying to tell everyone what to do...

Good memory management is certainly important, but I haven't seen 1600x1200 present problems with 8x MSAA on the vast majority of titles.

Perhaps if I was a member of a marketing team, or of some focus group wandering round internet forums to present some particular view of the facts then I might do that, but that's not generally how I would choose to spend my time.

With reference to your comment above I believe that your initial relevant post in this thread included the following statements -

"Heh. The advantages of CSAA far outweight the disadvantages of it."

"Nvidia users skip 8x MS because the advantages of using it are so minor and abstract that they just aren't that important."

Which looks to me much more like the sort of campaign that you're suggesting that I should pursue than anything that I've written in this thread

"8x Transparency SS is actually one of the things that makes 8x Multisampling so worthless"

Your statement didn't call transparency supersampling in conjunction with 8x MSAA worthless - it called 8x Multisampling as a whole worthless. I don't see any other way to read it.

Anyway since it's apparently not what you meant to write I won't belabour the point further.

Interestingly with custom filter AA you can also take just 4 samples, with the added benefit that the filter also improves the edges of the supersampled transparencies, whereas coverage samples give no benefits on these objects...

 

Chris ,

I understand your argument but my argument is why aren't their review doing more than just 4xMSAA benchmarking ?

Narrow-tent
Wide-Tent
Edge-detect

Like when your doing your review , do 16xCSAA benchmarking then.

 

The Tent modes aren't that interesting. Edge-detect is.

 

And the fact that tent filters blur the images to some extent unlike CSAA. However, edge detect on the other hand

 

hey twinkie ,

Performance vs Quality

Nvidia / ATI

Which is the best AA mode for 48XX and GTX2XX.

48XX = Edge-dent / MSAA ?
GTX2XX = CASS ?

 

I do bench at 16x CSAA and 16xQ. But I dont do comparative benchmarks to ATI hardware. ATI's and Nvidia's mode beyond 8x MS are not comparable. And the 8x MS mode is not really that great on Nvidia hardware. And I find the value of 8x multisampling to be extremely questionable when Nvidia users tend to focus on CSAA. Which is why most Nvidia users skip it. I really have no problem with reviewers who do comparative analysis to focus on edge detection, tents, and CSAA, But just doing them without talking about them in great detail is useless.

 

http://doomstar.awardspace.com/ATI_AA_modes.html

I created a rollover image comparison for ATI AA modes.
These image are taken from beyond3d article : AMD R600 Architecture and GPU Analysis.
http://www.beyond3d.com/content/reviews/16/15

What is Nvidia equivalent of ATI's 8x Edge Detect + Quality Adaptive AA mode. I know this mode only works in few games properly.

 

AT relatively unimpressed with the added value of the 9800 GTX+

 

The funny thing is that even you want to buy NVidia, the GTX 260 is only 0-20% faster than the 9800 GTX+ for almost twice the price.

ATI really messed up NV's product positioning. They probably lost 3/4 of their margin on G92/G92b, as I'm sure they had to lop off at least $60 from the chip price to have a $100 retail price cut in a product that's otherwise the same.

 

Does anyone know if it´s possible to use two 4850 CF for graphics and one 9800GTX for PhysX?

 

That's impossible.
In Vista the OS driver model doesn't allow graphics cards from different vendors to be running simultaneously on the same system.
In both XP and Vista i don't think Nvidia would ever bother to release a stand-alone CUDA/PhysX-only driver for their consumer GPU's (there's still the AGEIA PPU driver, but that line of software/hardware development is, in essence, dead).

 

Yep, part of Nvidia's problem right now is that GT200 didn't even deliver vs G92. All the synthetics show big gains but that didn't translate into actual games. What's going on there is still a mystery to me.

I wonder how they will react to the sort of arithmetic density and perf/mm^2 AMD has achieved. They can add more SM's per cluster but they still have the control overhead problem.

Maybe a ~300mm^2 256-bit GDDR5, 6-cluster part at 55nm with 1600Mhz+ clocks could put up a fight but I doubt they have anything like that in the pipeline. Nvidia was probably planning to fight RV770 with G92b....sure didn't turn out how they planned though.

 

It's not that much of a mystery. The only synthetics that show the gains bigger than the math or texture rate increase are the really crazy shaders and techniques that we know aren't in games. Looking at regular shaders, like those in D3DRightmark, the math and texturing is all that helped. G92b has 22% more texturing and setup speed and only 2% less math speed than the GTX 260. G92b has faster setup than the GTX 280, equal texturing speed, and only trails in math by 25%. Bandwidth is GT200's only big advantage.

As I said in another thread, NVidia has two problems with GT200. One is that they designed it to have lower peak perf/mm2/clock, possibly from underestimating AMD (hey, even I did) and thinking spending trannies on GPGPU would be mostly harmless. The other is that it's just not clocked as high as G9x chips.

If you exclude RV770, and scale all past chips from either manufacturer to 55nm, I think G92b is the chip that gives you the most performance per dollar (including all board costs, not just silicon) with today's workloads. It's a great chip. That's why it's so unbelievable that ATI blew past it, and at less than half the (shader) clock speed to boot.

 

...But with 6,25 times more shader processors. So is it 6,25 time faster than G92? No? Then what's unbelievable in it being a bit faster? By all accounts it should be! And not by a bit, but 2-3 times faster...
The only thing that is unbelievable about RV770 is the number of units packed into such size. Performance is pretty meh for what's packed into it. I believe that AMD could go with less ALUs and still hit the same performance targets.

 

RV770 is half the size of GT200, yet outperforms GTX 280 in some benchmarks, and you call the RV770 'meh'?

Compare G92b (9800 GTX+) to 4850. G92b has less ALUs, as you've noted, but they're clocked much higher (nearly 3x). If RV770 used such high clocks, there wouldn't be room for 800 ALUs. G92b has 64 texture units, vs. 40 for RV770. G92b texture units are clocked at ~700 mhz vs. 625 mhz for 4850. G92b uses higher memory clocks than 4850. G92b has more Z units than RV770. 4850 has far more double precision performance than GT200(!). With all these deficits (aside from ALUs), 4850 competes very well with G92b.

Yeah, I guess RV770 really is 'meh' afterall. LOL.

-FUDie