NV20 performance gains

[aniki]

What kind of performance gain are we looking at with the upcomming chip?
Please speculate if you have to

 

[JellyBaby]

Not certain but the lack of hype and noise on nv20 is deafening. Maybe I'm not hitting the right sites but there doesn't seem to be any Marketing Domination Campaign for this chip which I heard was being released this or next month...? Granted, why should nv spend the $$$ to fight zero competition but this is really unexpected...you'd think they might want to "flex" their victory a bit.

 

[Sunner]

Tehy'll probabaly launch a massive "surprise campaign" shortly before its release.
Take the world by storm

Was pretty much the same thing with NV15, there was hardly any confirmed info before its release, iXBT had some info that turned out to be mostly correct, and Creative hinted about the performance, but mostly there was no info at all until like a week before they were available.

 

[BFG10K]

I'd say 90 FPS at 1600 x 1200 x 32 in Quake 3 is not out of the question.

 

[Matt]

I'm just going to be boring and predict a 30% increase in performance
compared to the nv15.

Comments??

/Matt

There can only be one!

 

[RoboTECH]

BFG, that's a 50% increase over the GTS-Ultra

No way. I'll say 75 fps tops.

 

[Mooncalf]

I find it refreshing to not have a product hyped to death far in advance of release.

 

[jpprod]

90-100FPS in 1600x1200x32/HQ Quake III Demo001 easily.

I feel out of the current popular games, Quake III will be the one with most improvement. GeForce2 Ultra already does 60FPS, and it still isn't fillrate but memory bandwidth limited as overclocking tests indicate. NV20 will have (speculation, rumours) higher available memory bandwidth, more efficient rendering pipeline with hardware hidden surface removal plus (likely) larger texture caches. There's also a much, much faster T&L unit in the NV20. Assuming there aren't some still unknown factors to cap the performance, fillrate bound, relatively high-polycount game like Quake III will certainly see a big boost from NV16 -> NV20.

Here's how I, as a thick-headed Nvidia zealot, predict how NV20 will perform across the board:

• In vast majority of older games there will be no noticeable improvement, because they aren't fillrate bound and there really aren't that many polygons on the screen either. Examples of such games would be Unreal Tournament, Deus Ex, MDK2 and all them sh1tty EA-sports games. In these you'll only see a hefty high-res FSAA speed improvement due to the vastly incrased realworld fillrate.
• Relatively new games with modern fillrate-hogging engines, like Quake III and Serious Sam will see a very noticeable boost. Plus for the first time, I expect to be able to play high-resolution Q3 with FSAA on the NV20.
• In the very latest + upcoming titles utilizing DX8 there should be a much bigger difference, both in performance and visuals. NV20 T&L unit allegedly does much better to it's peak when polygons are small, and the geometry data is indexed.
• No doubt, I expect Nvidia to try and penentrate the professional OpenGL market with the NV20 as well. Professional applications should be close to the optimal situation for NV20's T&L unit - lots of small triangles, indexed and quite fixed geometry, most geometry modification operations done through high-level API calls. If T&L unit's accuracy isn't any worse than GeForce's (this does matter in the professional world), I believe NV20 will make a good pro chip as the drivers mature.

About the FSAA of the NV20.. I've been playing with the DX8 SDK for the few past days. Let me tell you, contrary to my expectations, the 9-sample multisampled FSAA looks sweet. I can't wait to see hardware support for this

Unconfirmed features I would most want to see in NV20:

1. Support for at least 16-tap anisotropic filtering, preferably even more. Radeon already does 16-tap anisotropic, and it looks noticeably better than GeForce (supports 8-tap) with high-detail mip-mapped textures in the distance.
2. No fillrate hit for trilinear/anisotropic. It isn't fun to turn off an excellent hardware feature like this just because it has a fillrate hit.
3. Single-pass (per pipeline) EMBM. EMBM might not be the most useful feature in video cards, and is overshadowed by much more flexible DOT3. However, it has it's uses. EMBM is a very cute way to fake reflections out of shiny surfaces, such as water or polished metal.
4. Hardware volumetric fog with variable density (defined by a function of displacement). DX8 doesn't directly support this, but I would find this feature very useful

edit: typos in the fusetalk code

 

[Remedy]

JP i can't believe they haven't made you an Elite member yet. :disgust:

 

[KarlHungus]

<< Let me tell you, contrary to my expectations, the 9-sample multisampled FSAA looks sweet. I can't wait to see hardware support for this. >>

Care to elaborate further as to why you didn't have high expectations?

Are you referring to a 9x FSAA method (3x3 -> 1), or a 1.5x FSAA method (3x3 -> 2x2)?

 

[jpprod]

KarlHungus,

The whole idea of multisampling is to re-use textel data between subsamples which hit the same polygon on the pixel's area. DirectX8 documents indicate that multisampling is done in a single pass, as triangles are rasterized. Because texture aliasing is handled in a different, not completely application-transparent process (trilinear or anisotropic mip mapping + filtering) compared to edge aliasing (traditionally blending subsamples), I expected there to be at least some compatibility issues. Or some unexpected issues to arise just because the technique's new to the API, and Microsoft's behind the whole mess.

Implementation of MS-FSAA isn't as straightforward and foolproof as SS-FSAA, but the output looks just as good. MS-FSAA on DX8 software reference rasterizer seems to run just fine and wihtout visual anomalies (though it's very, very SLOW) on all examples found in the SDK. This is a good sign, at least everything works on the API's side of things. With hardware implementation the speed should be much more tolerateable than the abysmal performance of SS-FSAA we know today.

There are two MS-FSAA modes available on the reference device in the demos, 4-sample and 9-sample one. These are 2x2 and 3x3 modes respectively, though I do not know specifically how subsamples are aligned (ordered grid, rotated grid or something even fancier). DirectX8 seems to support 0,2,3,4,5..16 samples to be used in MS-FSAA, I guess it's up to the hardware how to place them.

edit: Now I'm dropping verbs. This is why I'll never be an Elite member

 

[bluemax]

I speculate that it'll be over 2x to 3x faster! It'll support resolutions up to 3000x2000, 36x texture compression, 16x FSAA (which'll look like saran-wrap around your monitor) and will automatically double whatever system RAM you have.
Limitations: 2D so fuzzy you can't read text at 640x480, and limited to only 256-colour mode.

 

[JellyBaby]

<< 16x FSAA (which'll look like saran-wrap around your monitor) >>

bluemax, I may be a little slow today but that statement makes no sense to me.

At the resolutions nv20 is capable of, is fsaa even an issue anymore? As far as jaggy reducation goes running at 1600+ ought to address that problem.

 

[MSNY]

The NV20 is in my sights as my next upgrade from my GeForce DDR.

My main concern is no so much all the great features, I'm sure they will be there for the most part. I'm concerned with $$$ !

 

[AudiPorsche]

im guessing that it will reach at least $800 for a nv20 ultra type card.

 

[Sunner]

I doubt it.
nVidia arent stupid, they arent going to give the market away by ridiculous prices.

 

[TravisBickle]

there is not much point hyping or advertising the nv20 as hardly anyone will be able to buy it since it must be faster than a GTS ultra.
Rolls Royce and Lamborghini do quite well without plastering my Sunday papers with ads

 

[KarlHungus]

jpprod -

Cool, thanks for the explanation. Got any links to whitepapers on the subject?

 

[Packet]

I would not be surprised if the NV20 is as fast as jpprod suggests. Being that its a whole new chip, and not another high clocked Geforce, its obviously going to have some speed gains just in how they engineered the chip. The drivers will also already be quite mature, (the latest leaked drivers have hints of NV20 compatability in the inf files) and Nvidia has already shown to be kings in terms of driver updates.

If you look at the drivers from when the Geforce first was released to now, there is not only substantial speed increases, but features as well. (Texture compression, FSAA even though right now its not as nice as 3dfx, etc)

I am guessing, that there will also be a bit of scaling with CPUs when T&amp;L is not involved (Much like what happened when the V2 first came out, you need a faster CPU to take full advantage over it, etc, etc).

 

[jpprod]

Here's the same page as is provided in Microsoft DX8 SDK. It isn't very informative, but shows how easy it is to implement support for it once hardware has support - it's a simple on/off switch just as SS FSAA. It also indicates that developer has control over operations which are multisampled - there's no point in, for example, apply multisampling on particle effects.

Here's a bit about multisample rendering in general of which MS FSAA's just an application. In MS FSAA selectivity factor is whether subsamples in a pixel's area are using the same texture.

 

[DaveB3D]

What is described at RC is a form of super-sampling via multisample combining, not multi-sampling anti-aliasing. NV20 uses multi-sampling anti-aliasing. There will be an article posted on the subject this week by yours truly.

 

[KarlHungus]

Hey Dave, post the URL in this thread when you get the article ready. That is if you don't mind, of course.

 

[DaveB3D]

sure thing...

 

[aniki]

How well will the T&amp;L help with the lower to mid end cpu's. Will we need a 1.5 ghz to enjoy large performance increases?

 

[Yoshi]

I'm sure it will be very fast, but I have not found anything that dogs down my GeForce DDR based card so I will be holding off.