ATI's thoughts on Shader Model 3.0

[tranCendenZ]

With all the talk of SM2.0b vs SM3.0, what does ATI's Richard Huddy have to say on the subject?


Heh, check out this leaked powerpoint presentation:

http://pwp.netcabo.pt/0239863201/Save_The_Nanosecond.ppt

Look at the notes on Page 31:
* It’s generally agreed that SM3.0 requires a move to 32 bit floats (some still think it’s not)
* But - The spec requires it...
*Blend etc will be on NV40 and on our own SM3.0 parts
*Steer people away from flow control in ps3.0 because we expect it to hurt badly. [Also it’s the main extra feature on NV40 vs R420 so let’s discourage people from using it until R5xx shows up with decent performance...]

This same powerpoint presentation is available in PDF format directly from ATI without notes at:
http://www.ati.com/developer/gdc/Huddy_SaveTheNanosecond.pdf
Same content though mentioning Dynamic Flow control causing performance issues even on the one on ATI's site.

Funny

 

[@trapine]

The funny thing with instancing is ALL Radeon DX9 Class cards support it.And from the looks of it get a better performance boost than the NV40 chips?

Yesterday we published details on FarCry 1.2, part of that editorial focused on the ability to enable Instancing on the Radeon X800 within FarCry. Since publishing the details we have been working with www.tommti-systems.de to get their Nvidia Instancing demo working on the X800. With very little trouble (only a change to the program which removed a SM3.0 check) Tommti provided a fixed exe and the demo was up and running.

To recap for those of you who dont know or have forgotten what Instancing is... Instancing is a technique where large numbers of identical objects can be grouped together and processed as a batch, greatly reducing the work required by your system. Example real gameplay uses are trees and grass in FarCry and in the Tommti Demo Asteroids or Ships. With traditional rendering each object has to be processed seperately.

The demo works in a very simple but effective way, a scene is rendered where the same object (asteroids or ships) is displayed a large number of times. The number of objects is user definable however asteroids is set to 16,000 by default and ships to 1. With regards to running the demo there are 2 modes which can be chosen, Instancing on and Instancing off. The program monitors peformance and reports the current FPS to the user. It is also possible to move the camera throughout the scene if you require to do so.

With the patch up and running (Using the same 8.041 ATI drivers used for the Farcry patch) we monitored performance.

For reference Tommti's 6800GT clocked at Ultra Speeds achieved up to 3fps with instancing turned off and up to 40fps with instancing enabled (Shader Model 3 support).

Our X800XTPE achieved 4.5fps with instancing off and 48fps with instancing enabled (SM2.0b).

Nvidia have been more than enthusiastic about the benefits of Instancing support in their NV40, we can see they were absolutely correct to feel this way. Implemented correctly its a great feature and not only does the Radeon X800 support it and the performance benefit is equally large, other DX9 class ATI cards also support this feature.

Lets hope more developers follow Crytek's lead and implement this feature for the NV40 and all DX9 class Radeons asap.

 

[tranCendenZ]

The funny thing with instancing is ALL Radeon DX9 Class cards support it.And from the looks of it get a better performance boost than the NV40 chips?

Instancing is only a part of SM3.0, I can list a whole page of stuff that the X800 can't do that the 6800 can with sm3.0.

Do you find it good that ATI is discouraging developers from using more advanced graphical features because the X800 is behind in technology?

 

[^eMpTy^]

The funny thing with instancing is ALL Radeon DX9 Class cards support it.And from the looks of it get a better performance boost than the NV40 chips?

If it's supported then why are people having to hack stuff up to get demos running to prove it? If this is true I wish we could get someone with some authority to comment on it...I don't trust these demos that people make...but this is a pretty important question that should be answered.

Personally I'm guessing this is ATi fans just trying to downplay the new features of PS3 by making demos showing that ATi can do it too...which is understandable...but hardly trustworthy...like I said, we need some info from CryTek on the subject...hell I'd even take an OFFICIAL word from ATi...but nothing like that is around yet...

 

[@trapine]

Instancing is only a part of SM3.0.

Do you find it good that ATI is discouraging developers from using more advanced graphical features because the X800 is behind in technology?

The thing i find most funny is when Nvidia first launched 6800.SM3 was its big claim to fame.And we were hearing how it was gonna hurt ATI and X800?But by most of current accounts ATI's way of going about it seems to be the better performance winner?So i give ATI cudos for staying the course with 24BitP and SM2B.And then move over to SM3 at the end of the year when Clockspeeds have increased to match the performance hit!

 

[^eMpTy^]

The thing i find most funny is when Nvidia first launched 6800.SM3 was its big claim to fame.And we were hearing how it was gonna hurt ATI and X800?But by most of current accounts ATI's way of going about it seems to be the better performance winner?So i give ATI cudos for staying the course with 24BitP and SM2B.And then move over to SM3 at the end of the year when Clockspeeds have increased to match the performance hit!

lol

Yeah Nvidia was all freaking out trying to beat ATi...when all they had to do is sit by and watch ATi shoot themselves in the foot with the non-existant XTPE and shitty OpenGL drivers...

Now THAT'S funny.

 

[tranCendenZ]

The thing i find most funny is when Nvidia first launched 6800.SM3 was its big claim to fame.And we were hearing how it was gonna hurt ATI and X800?

Apparently even ATI themselves believes this as evidenced by their own powerpoint presentation above.

But by most of current accounts ATI's way of going about it seems to be the better performance winner? So i give ATI cudos for staying the course with 24BitP and SM2B.

ATI can go from a marginal performance victory to a big performance defecit if devs actually utilize SM3.0, as evidenced by ATI's own presentation.

Here are some of the areas where SM2.0b is behind SM3.0:

Dependent Texture Limit
2.0b = 4
3.0 = No Limit

Position Register
2.0b = none
3.0 = Yes

Executed Instructions
2.0b = 512
3.0 = 65536

Interpolated Registers
2.0b = 2+8
3.0 = 10

Intstruction Predication
2.0b = none
3.0 = Yes

Indexed Input Registers
2.0b = none
3.0 = yes

Constant Registers
2.0b = 32
3.0 = 224

Arbitrary Swizzling
2.0b = none
3.0 = yes

Gradient Instructions
2.0b = none
3.0 = yes

Loop Count Register
2.0b = none
3.0 = yes

Face Register (2-sided lighting)
2.0b = none
3.0 = yes

Dynamic Flow Control Depth
2.0b = none
3.0 = 24

Minimum Required Full Precision Shader Capabilities
2.0b = FP24 (96-bit)
3.0 = FP32 (128-bit)

And then move over to SM3 at the end of the year when Clockspeeds have increased to match the performance hit!

ATI is behind in tech, it was a misstep this round. They are doing a fine job of spinning, though.

 

[Lord of Shadows]

bah thats old news =P

 

[tranCendenZ]

bah thats old news =P

Heh I never read it before.

 

[@trapine]

bah thats old news =P

I also agree these slides got done to death on here ages ago.Just more people trying to Hype Smoothmarketing3

 

[^eMpTy^]

I also agree these slides got done to death on here ages ago.Just more people trying to Hype Smoothmarketing3

And you're trying to hype ATi's counter-smoothmarketing...funny part is, Nvidia doesn't even need PS3.0 to be the more attractive card...

 

[tranCendenZ]

And you're trying to hype ATi's counter-smoothmarketing...funny part is, Nvidia doesn't even need PS3.0 to be the more attractive card...

truth

 

[@trapine]

Nvidia doesn't even need PS3.0

Well clearly they think they do lmao because they did one hell of a holler about all the games that are coming out that will use it.And it was gonna increase performance by 20% and leave X800XTPE standing Well that just aint happened
And to boot ALL Radeon DX9 cards can benefit From the performance increase!

 

[tranCendenZ]

Well clearly they think they do lmao because they did one hell of a holler about all the games that are coming out that will use it.And it was gonna increase performance by 20% and leave X800XTPE standing Well that just aint happened

yet

The 5900 Ultra was as fast as the 9800PRO for about six months, until FarCry came out.

5900 Ultra owners now have to upgrade to get even decent performance out of intensive SM2.0 titles while 9800PRO owners have a choice. The same may happen with X800 and intensive SM3.0 titles.

What happens when SM3.0 titles start to use Dynamic Flow Control and performance "hurts" for X800 as ATI's Richard Huddy puts it?

And it was gonna increase performance by 20% and leave X800XTPE standing

Sounds like you are thinking of the $399 6800GT versus the $499 X800XTPE in Doom3

 

[^eMpTy^]

Well clearly they think they do lmao because they did one hell of a holler about all the games that are coming out that will use it.And it was gonna increase performance by 20% and leave X800XTPE standing Well that just aint happened

I'll say it again. The XTPE is completely unavailable. Even when it is available it barely beats the 6800U if AT ALL in every game. Nvidia clearly wants to pull out all the stops to get thier performance crown back...which they have now done in my opinion.

Instead of bitching about overhyped features, why don't you honestly recommend what card you think people should buy right now...and I dare you to say "XTPE".

 

[@trapine]

Noone is disputing That 6800GT isnt the best current price for buck card empty But some people in here are clearly starting to get a tad bit over excited because of the Doom3 benchies. And suddenly think there 6800GT'S will spank X800XTPE in any game.And that is just clearly FUD and not the case at all.X800XTPE is starting to trickle on to the market now.And when it does hit volume Its gonna be intresting to see.Some real numbers coming in
Whether you like it or not SM3 is just a marketing ploy at the moment But yes it will be useful eventually.Its abit like the 64Bit functionality of the A64 Only you can utilize 64Bit on A64 Now

 

[^eMpTy^]

Noone is disputing That 6800GT isnt the best current price for buck card empty But some people in here are clearly starting to get a tad bit over excited because of the Doom3 benchies. And suddenly think there 6800GT'S will spank X800XTPE in any game.And that is just clearly FUD and not the case at all.X800XTPE is starting to trickle on to the market now.And when it does hit volume Its gonna be intresting to see.Some real numbers coming in
Whether you like it or not SM3 is just a marketing ploy at the moment But yes it will be useful eventually.Its abit like the 64Bit functionality of the A64 Only you can utilize 64Bit on A64 Now

I have no delusions about PS3...you seem to have some about XTPE availability though...

I don't recall anyone saying that the GT will spank the XTPE in any game, clearly that's incorrect, the XTPE is a fast ass card.

I don't really care about whether or not SM3 is a marketing ploy. The simple fact is, ploy or not, SM3 is in the 6800 and it's not in the X800. ATi can say it's unimportant all they want but we all know it will become important the moment they support it.

And if you're not trying to dispute the fact that the ONLY card to buy right now is the GT, then why are you incessantly bashing Nvidia? Obviously Nvidia is going to hype thier new feature...it's called MARKETING. So why are you so stuck on it?

 

[Jaffrin]

Yeah, OK, the X800 XTPE is out performing the 6800 Ultra in reviews
(Depending on who benchmarked them)
But considering that the ATi shortcut in filltering is one of the reason why...

nVidia would be getting better fps in certain reviews right now if it was taking
the same shortcuts.

I have read countless reviews of both ATi and nVidia's cards and they differ
as the days go on.
@trapine it is obvious you are simply blinded by your bias.

You can argue that the Shortcuts are trivial,
but any form of compression which is intended to
shorten the amount of time it will take to render an object will make the card faster.
I am not saying it is bad, I am simply saying that some of the fps that ATi is claiming
in the 4aa 16xaf 1600x1200 benchmarks would likely drop.

 

[R1ckCa1n]

Is it me or has the video forum turned into attack anything related to ATI and hype up everything about Nvidia; even if it's at partial precision?? If it's "intrepreted" bad for ATI, it will be posted here by several people and relentlessly persued. That's OK, but it does shine an incorrect light on ATI's current products. Just my 2 cents.

 

[tranCendenZ]

Yeah, OK, the X800 XTPE is out performing the 6800 Ultra in reviews
(Depending on who benchmarked them)
But considering that the ATi shortcut in filltering is one of the reason why...

nVidia would be getting better fps in certain reviews right now if it was taking
the same shortcuts.

The funny thing is that Nvidia has similar shortcuts but some reviewers, such as DriverHeaven, are choosing not to enable them, making extremely biased benchmarks.

 

[^eMpTy^]

Is it me or has the video forum turned into attack anything related to ATI and hype up everything about Nvidia; even if it's at partial precision?? If it's "intrepreted" bad for ATI, it will be posted here by several people and relentlessly persued. That's OK, but it does shine an incorrect light on ATI's current products. Just my 2 cents.

After you've already beaten the partial precision question to death you still choose to bring it up.

The fact is right now Nvidia is doing everything right and ATi is doing everything wrong. So if you go to some ATi biased sites and come back to this site with a bunch of ATi marketing bs and try to convince people it's worth while, you're going to get beat down because the people on this forum are smart enough to know the difference.

 

[tranCendenZ]

Is it me or has the video forum turned into attack anything related to ATI and hype up everything about Nvidia; even if it's at partial precision?? If it's "intrepreted" bad for ATI, it will be posted here by several people and relentlessly persued. That's OK, but it does shine an incorrect light on ATI's current products. Just my 2 cents.

It's difficult not to "hype up" Nvidia. They have made the right choice at nearly every branch this round, from having the most advanced featureset (SM3.0/FP16 blending/Ultrashadow II) to the most flexible shader architecture for developers (FP32 for highest quality, FP16 for a 5-15% performance boost with no IQ loss on many shaders) to making the most options available to end users (Both Full & optimized Trilinear filtering, SuperSampling AA) to excellent performance both in Direct3D and OpenGL.

That's not to say ATI cards are bad or that threads like this are an "attack" on ATI, it's just to put into perspective ATI's choices and the marketing that goes along with them. ATI's cards are excellent performers this round, but they come up short in features and end user customization. They might be an okay choice for someone who plans to buy a videocard 6 months from now anyway, but it makes it difficult to recommend them when Nvidia cards give similar or better performance with a much more robust and future-proof featureset.

All I can say is that if you are taking what ATI's marketing is saying about Shader Model 3.0 and it being "worthless" at face value, or that SM2.0b is just as good, take a look at this powerpoint presentation by ATI. It shows that even ATI believes there are some SM3.0 features that will make the X800 have serious performance issues, and ATI will be doing their best to make sure they don't make it into games until ATI's SM3.0 cards come out. But when ATI SM3.0 cards do come out, X800 owners may find themselves in the same positions FX5900 owners are finding themselves; cards that were fast when they first came out, but unable to keep up with the new games due to lacking what will be key DX9 features.

Again though, it might be 6 months before we see these features implemented in games with great intensity, but the 6800GT/6800U seems to have the best potential for lasting the longest out of all the next-gen cards. They may become another ti4200 or 9700PRO, cards that last several generations while still offering an excellent gaming experience.

 

[fallguy]

And the last time I checked the X800 line sucks because ATi didn't bother to write a decent OpenGL driver...so what?

The fact is right now Nvidia is doing everything right and ATi is doing everything wrong. So if you go to some ATi biased sites and come back to this site with a bunch of ATi marketing bs and try to convince people it's worth while, you're going to get beat down because the people on this forum are smart enough to know the difference.

First off, this has been posted here several times already. Secondly, how can you call anyone else bias with comments like these? The XT/PE is faster in the majority of games, and especially in Farcry which has been NV's baby as of late. Trying to show it off whenever it can. You can say then are doing everything wrong, but the benchmarks do not back up your statement. Only in Doom3 is the 6800's (so far) getting a large lead. Saying the X800 line sucks is about as ignorant as you can get. Talk about being smart enough.

Dont take offense to this, its not ment as an insult towards you, but you're going on ignore. 50 posts per day spamming the forums while being obviously very pro-NV is annoying. Posting old topics is annoying. Seems to me you were registered here once already. I own a 6800GT before you try to call me some sort of fanboy.

 

[R1ckCa1n]

After you've already beaten the partial precision question to death you still choose to bring it up.

The fact is right now Nvidia is doing everything right and ATi is doing everything wrong. So if you go to some ATi biased sites and come back to this site with a bunch of ATi marketing bs and try to convince people it's worth while, you're going to get beat down because the people on this forum are smart enough to know the difference.

Exactly what is ATI doing wrong? Do they lower precision to make up for performance gaps? Do they hype SM3.0 when they don't do it per spec and get beat with a card running the "old technology" SM2.0b? The only thing the NV40 really has going for right now is Doom3. This is ONE game.

The sad part about this is ATI will have to sink to the lows of Nvidia just to "Get in the Game". I think we will see evidence of this on the next ATI driver release.

Sadly you comment "you're going to get beat down because the people on this forum are smart enough to know the difference" is more about bias than fact.

 

[HeavenX]

The new ATI cards are not even worth that much money, they dont offer anything new after the radeon 9700 core was created. Nvidia cards today is the new technology and the best to go if you want to spend money on a powerfull video card that will give you best fps and best realism. I have 9800 pro right now and I choosed it over nvidia's nv30 core and I am not biased. I know which video card is better to buy now and that is the nvidia 6800.

 

[fallguy]

The new ATI cards are not even worth that much money, they dont offer anything new after the radeon 9700 core was created.

Wrong. While its your choice to get which card you want, you may want to do more research before doing so.

 

[R1ckCa1n]

The new ATI cards are not even worth that much money, they dont offer anything new after the radeon 9700 core was created. Nvidia cards today is the new technology and the best to go if you want to spend money on a powerfull video card that will give you best fps and best realism. I have 9800 pro right now and I choosed it over nvidia's nv30 core and I am not biased. I know which video card is better to buy now and that is the nvidia 6800.

See, this is exactly what this forum (acutally a few 50 plus a day posters) is creating. I had a 9800ProXT and went to the X800XT-PE and saw my performace DOUBLE. How do you say it is not worth the money? They offer a ton of new features over the 9700 core. Because you buy into all the hype of this forum (acutally a few 50 plus a day posters), you are blinded by the fact this old core is winning almost every benchmark?

Yes the NV40 was a HUGE product update for Nvidia, but there's more smoke and mirrors than fact to the product. You'd think people would have learned by the FX series....

 

[tranCendenZ]

Exactly what is ATI doing wrong? Do they lower precision to make up for performance gaps?

No, they just use a lower FP24 precision to start with instead of giving the developer the power to choose what precision they want to use. FP32 for highest quality, FP16 for a 5-15% performance boost for shaders that dont need FP24 or FP32. With Nvidia cards, developers have the option to specify whichever precision they want in their code, giving the best combination of performance and IQ.

Do they hype SM3.0 when they don't do it per spec

Nope, but Nvidia hyped SM3.0 and is doing it per spec. I'd like you to point out where they aren't. If you are about to say partial precision, think again, because partial precision is in the SM3.0 spec.

and get beat with a card running the "old technology" SM2.0b?

Let's wait and see til multiple pro review sites take this on. It is possible that SM2.0B path in FarCry is faster than the beta SM3.0 path currently, as the beta SM3.0 path doesn't use much SM3.0, and also operates at a higher precision for some shaders. IUn addition, Crytek they have stated that they are going to add more SM3.0 features to the 1.3 patch - which makes sense being that in the 1.2 patch the SM3.0 path is in beta nad can't even be enabled without using developer mode.

That being said, in the future, as ATI themselves points out, there are features of SM3.0 that will cause serious performance issues for their sm2.0b cards. We have seen many reviewers totally screw up benchmarks, such as the recent DriverHeaven review which enabled trilinear optimizations on ATI cards but disabled them on Nvidia cards, which makes no sense and gives biased results. Therefore it will be best to evaluate this once we have a patch that runs well without visual glitches (ie textures disappearing on x800 cards) on both cards.

The only thing the NV40 really has going for right now is Doom3. This is ONE game.

NV40 has Doom3 plus a slew of features that ATI cards lack. It also gives you the best performance per dollar with the 6800GT features aside. Even if you didn't use those features, Nvidia delivers a 16 pipe card at $399 versus ATI's 12-pipe card, and their OpenGL implementation will be ahead not only in Doom3, but also Quake4, RTCW2, and other upcoming OpenGL games. Nvidia's featureset and performance makes their cards the most future proof at this juncture, and the best buy.

The sad part about this is ATI will have to sink to the lows of Nvidia just to "Get in the Game". I think we will see evidence of this on the next ATI driver release.

If by sink low you mean deliver the best price/performance/features bang-for-the-buck video card, you are right. Except its going to be impossible for ATI to meet Nvidia's features this round.

 

[Kevin Lowe]

Do they lower precision to make up for performance gaps?

Do you actually understand the concept of precision? Have you ever written a line of code in your life?

I'm assuming no, so let's have a quick lesson for you.

Floating point precision is most important when you're performing multiple operations and have to preserve tiny portions of the number from one operation to the next, or when you're working with incredibly large numbers that can't be represented with fewer bits. Adding precision is great because your results can be more accurate; however, you pay a cost in memory bandwidth (and possibly processing time, depending on how the chip is designed) to move the larger numbers around, and store the intermediate results of each successive operation.

So, if those tiny portions of the number don't actually change the result within the domain of potential values then there's no point in actually carrying them around from one stage to the next. By carefully analyzing each part of a program, you can determine when it's acceptible to reduce precision to optimize performance.

Eventually, with the number of operations taking place, the tiny rounding errors will add up to a point where they have a perceptible impact upon the end result. In that case, you step the precision up, and the rounding errors go away (until you add operations, in which case you eventually need even more precision.) However, if your programs don't perform that many operations, the accrued rounding errors will never be noticed, and the overall result will be unaffected.

To conclude, if your system were eventually rounding everything to an integer before writing out the final value (for instance, a frame buffer), then any routine where reducing precision resulted in a change of less than 0.5 in the final value (pre-rounding) should be optimized for performance instead.

That's why it doesn't matter when developers use reduced precision for certain shaders - they've already carefully analyzed the results and determined that it doesn't make a difference either way.

 

[^eMpTy^]

Exactly what is ATI doing wrong? Do they lower precision to make up for performance gaps? Do they hype SM3.0 when they don't do it per spec and get beat with a card running the "old technology" SM2.0b? The only thing the NV40 really has going for right now is Doom3. This is ONE game.

The sad part about this is ATI will have to sink to the lows of Nvidia just to "Get in the Game". I think we will see evidence of this on the next ATI driver release.

Sadly you comment "you're going to get beat down because the people on this forum are smart enough to know the difference" is more about bias than fact.

The only thing sad here is your failure to comprehend that partial precision is not a cheat, it's SUPPOSED TO WORK THAT WAY. Ask ATi what their next architecture will support...FP24 or FP16/FP32.

As to what ATi is doing wrong: Not providing a down-clocked 16 pipe part to compete with the GT...not allowing thier optimizations to be disabled on a $500 card for better image quality...sitting on thier ass for 2 years with the r300 core instead of re-writing thier opengl driver...whining about partial precisions when they know there is nothing wrong with it just like everyone else knows...sticking with the same damn core for 3 generations instead of developing something that can compete with the feature set of the 6800 series...not providing a decent linux driver...not providing decent multi-monitor support...not getting the XTPE to market (though this isn't really their fault)...

 

[DryFire]

Ok before i start i'd like teo say i think the 6800gt right now is teh best bang/buck card.

Nope, but Nvidia hyped SM3.0 and is doing it per spec. I'd like you to point out where they aren't. If you are about to say partial precision, think again, because partial precision is in the SM3.0 spec.

how can you prove this? We have yet to see any game use all that SM3.0 has to offer.

SM3.0 offer a lot but odds are the people rushing out to buy x800pro xt-pe's and 6800Ultra's will be buying new card before we even see it in most games.

 

[ZenOps]

Hey, I'd be willing to be that there is almost an exact same memo from a Nvidia rep floating around somewhere that is being sent to developers.

32-bitFP can be brutally taxing on speeds, its not like going from 16 to 32INT. I don't think its in Nvidia or ATi's best interests to be hyping up 32-bitFP at all (stick to 16 and 24FP as is for now)

 

[^eMpTy^]

Wrong. While its your choice to get which card you want, you may want to do more research before doing so.

I agree...they use approximately the same core...but they have a hell of a lot more performance to offer in comparison to ATi's previous generation...

 

[tranCendenZ]

Hey, I'd be willing to be that there is almost an exact same memo from a Nvidia rep floating around somewhere that is being sent to developers.

32-bitFP can be brutally taxing on speeds, its not like going from 16 to 32INT. I don't think its in Nvidia or ATi's best interests to be hyping up 32-bitFP at all (stick to 16 and 24FP as is for now)

That's incorrect. In Farcry, the difference between FP16 and FP32 is approximately 5-15% depending on the scene. It has already been benchmarked in full FP16 versus full FP32. Full FP32 is about 5%-10% slower most of the time, translating to 3fps to 6fps slower. In 3dmark03 Game Test 4, the difference between x800 FP24 and 6800 full FP32 is about 15% or so. Your perception of FP32 performance on the 6800 is incorrect.

So why use FP16 if the performance difference is only 5-15%? Because there is zero IQ loss when used properly and every fps helps. And as games get more and more complex and more and more shaders are used, like in Unreal 3, partial precision will become even more important

 

[kcthebrewer]

Lets wait for Battle for MIddle Earth and look at their performance numbers. It is supposed to be the first game written to use SM3.0 instead of an add on. I believe it will give the full story of performance SM2.0x vs SM3.0.

 

[burningrave101]

Hey, I'd be willing to be that there is almost an exact same memo from a Nvidia rep floating around somewhere that is being sent to developers.

32-bitFP can be brutally taxing on speeds, its not like going from 16 to 32INT. I don't think its in Nvidia or ATi's best interests to be hyping up 32-bitFP at all (stick to 16 and 24FP as is for now)

What do you mean it can be brutally taxing on speeds? I've heard reports that the 6800's run FP32 nearly as fast as FP16 and its more efficient because of the higher precision level. The only time they use FP16 is when FP32 brings no real benefit and there is no IQ loss from doing so.

how can you prove this? We have yet to see any game use all that SM3.0 has to offer.

SM3.0 offer a lot but odds are the people rushing out to buy x800pro xt-pe's and 6800Ultra's will be buying new card before we even see it in most games.

People are going to be buying new cards again before the end of this year? SM 3.0 will be in over a dozen games just THIS year. Then we have all of next year and maybe longer. SM 4.0 wont be on the scene until the R600 and NV60 cores debut with Longhorn and DirectX 10. We've only seen SM 3.0 implemented into ONE game and Crytek did a rather poor job of taking advantage of what SM 3.0 can really do.

 

[R1ckCa1n]

That's incorrect. In Farcry, the difference between FP16 and FP32 is approximately 5-15% depending on the scene. It has already been benchmarked in full FP16 versus full FP32. Full FP32 is about 5%-10% slower most of the time, translating to 3fps to 6fps slower. In 3dmark03 Game Test 4, the difference between x800 FP24 and 6800 full FP32 is about 15% or so. Your perception of FP32 performance on the 6800 is incorrect.

So why use FP16 if the performance difference is only 5-15%? Because there is zero IQ loss when used properly and every fps helps. And as games get more and more complex and more and more shaders are used, like in Unreal 3, partial precision will become even more important

I did some research on M$'s website and see nothing abouth FP16/FP32 blending as part of DX9.0c/SM3.0 spec; all I see is minimum FP32. Please provide a link from M$ or any other reliable source stating otherwise.

 

[Taurus]

The only thing sad here is your failure to comprehend that partial precision is not a cheat, it's SUPPOSED TO WORK THAT WAY. Ask ATi what their next architecture will support...FP24 or FP16/FP32.

As to what ATi is doing wrong: Not providing a down-clocked 16 pipe part to compete with the GT...not allowing thier optimizations to be disabled on a $500 card for better image quality...sitting on thier ass for 2 years with the r300 core instead of re-writing thier opengl driver...whining about partial precisions when they know there is nothing wrong with it just like everyone else knows...sticking with the same damn core for 3 generations instead of developing something that can compete with the feature set of the 6800 series...not providing a decent linux driver...not providing decent multi-monitor support...not getting the XTPE to market (though this isn't really their fault)...

I have to agree with you on this buddy. I believe your 100% right. Nvidia is ahead in terms of technology.

 

[R1ckCa1n]

We've only seen SM 3.0 implemented into ONE game and Crytek did a rather poor job of taking advantage of what SM 3.0 can really do.

That is quite the change from all the hype you have produced over the 1.2 patch and how great SM3.0 will be for Far Cry as far as IQ and speed.

 

[tranCendenZ]

I did some research on M$'s website and see nothing abouth FP16/FP32 blending as part of DX9.0c spec; all I see is minimum FP32. Please provide a link from M$ or any other reliable source stating otherwise.

Whoa there FP16/FP32 blending is different than FP16/FP32 precision.

FP16/FP32 blending is used for OpenEXR HDR; Nvidia currently supports the FP16 format in the 6800, ATI does not support it with the x800 at all.

As for partial precision, this is the quickest link I could come up with that details the various shader models and their features, it goes into great depth from a reliable source:
http://www.beyond3d.com/previews/nvidia/nv40/index.php?p=5

When examining the charts, note that X800 cards fall under the "2.0B" category and 6800 cards fall under the "3.0" category.

From the link:
"Another element to note is that high precision mode for Shader 2.0 was set for FP24, however Pixel Shader 3.0’s high precision now comes in at FP32 – FP24 and below now becomes partial precision in the Pixel Shader 3.0 model."