S3's DX9 Deferred Renderer

[Brent]

http://www.s3graphics.com/index.html

http://www.savagenews.com/?show=2726

whitepapers there, i haven't looked at them yet

 

[Ailuros]

Advanced Deferred Rendering
Hierachical Z
In addition to conventional z buffering, where z values are computed and compared at the
pixel level, Columbia supports another coarser z buffering scheme in advance of the
standard z buffering stage. An incorporated zero-cycle clear and optional deferred
rendering (two pass) mode further optimizes color/z-buffer memory traffic and
pixelperformance. Deferred rendering virtually eliminates opaque (redundant) back-to-front
overdraw and saves also precious conventional z write traffic.

Sounds more like an IMR to me that uses a combination of DF and Hierarchical Z through a traditional pipeline. The optional DF note above reminds me of SA's notes on combinations of application driven deferred rendering combined with Hierarchical Z.

edit:

Is there any AA enabled at all in the pics in the "SilkySmooth" pdf, or have my eyes gone bad?

 

[Brent]

Advanced Deferred Rendering
Hierachical Z
In addition to conventional z buffering, where z values are computed and compared at the
pixel level, Columbia supports another coarser z buffering scheme in advance of the
standard z buffering stage. An incorporated zero-cycle clear and optional deferred
rendering (two pass) mode further optimizes color/z-buffer memory traffic and
pixelperformance. Deferred rendering virtually eliminates opaque (redundant) back-to-front
overdraw and saves also precious conventional z write traffic.

Sounds more like an IMR to me that uses a combination of DF and Hierarchical Z through a traditional pipeline. The optional DF note above reminds me of SA's notes on combinations of application driven deferred rendering combined with Hierarchical Z.

edit:

Is there any AA enabled at all in the pics in the "SilkySmooth" pdf, or have my eyes gone bad?

I think you are right, its an IMR, since it is having to deal with the Z-buffer

In the pics I can see the difference with AA enabled

It does support some impresive looking things on paper

looks like on par with the R300, that support for 16X SuperSampling and MultiSampling AA looks interesting?

hmm....

 

[Althornin]

looks feature rich, but a few too many marketing terms right now.
reminds me of the Xabre with all the marketing speak.

 

[Ailuros]

Someone zoomed in the images on another board and it makes AA more apparent.

I didn't see any reference for an amount of samples concerning SSAA, just MSAA. Lucky guestimate is the algorithm adaptive?

 

[JF_Aidan_Pryde]

Looks almost set to flop. That's how it feels anyway..

 

[Nebuchadnezzar]

It seems to be exatcly the same as R300...

But now:

1. Is it 0.13µ / 0.15µ ?
2. 128 bit / 256 bit bus ?

 

[Kristof]

Advanced Deferred Rendering
Hierachical Z
In addition to conventional z buffering, where z values are computed and compared at the
pixel level, Columbia supports another coarser z buffering scheme in advance of the
standard z buffering stage. An incorporated zero-cycle clear and optional deferred
rendering (two pass) mode further optimizes color/z-buffer memory traffic and
pixelperformance. Deferred rendering virtually eliminates opaque (redundant) back-to-front
overdraw and saves also precious conventional z write traffic.

"Optional" sounds like it does not always work. Two-pass sounds like they might be doing this :

1) Render geometry with depth only enabled hence building a Z buffer for the complete scene. Store state and geometry data at the same time (driver level).
2) Render the stored data again this time enabling the Pixel Shader. Since you already have the Z information for the whole scene from pass 1 your early Z should work independent of the render order.

This would effectively remove the EarlyZ dependence on the render order but at which cost ? DoomIII style rendering at the driver level.

They need to store all the scene data but where is this storage pre-VS or post-VS... if its pre-VS it will be a lot of data... and if the driver is doing this its a lot of overhead work (CPU bottleneck).

What about bandwidth... geometry flows into the design twice possibly meaning VS has to work twice as well. The building of the depth information costs pixel shader clocks. They still use a depth buffer that needs to be accessed in external memory.

Dunno just guessing. This all does not look like final material, me thinks someone is building the site but accidentially updated the real thing...

And "silky smooth" ? Thats what my washing powder claims as well...

 

[Dreamer]

When I saw the page I though that I has been hacked or something... it is soooo unproffesional and unfinished. Also, as Kristof pointed, some things sound if not funny then not really serious.

I have a riddle for all of you: What's that thingy that S3 announces as tech and it doesn't work as speacified?
A: everything

 

[Randell]

anyway who's going to write the drivers now that OpenGL guy and co are at ATI?

 

[mboeller]

Is this for real?

For me the PDF's look like an fake.
No logo's; nothing at all indicates that this PDF's are for real. All the "information" given is generic DX9 information + HDTV-information + R300 (24bit info).
So IMHO this is an fake. Seem's like the S3-homepage was hacked.

 

[Ailuros]

If that truly is another hoax, the guy who went through all this trouble could have been more resourceful and humorous.

Silky Smooth(tm) should then be accompanied with Fuzzy Logic(tm).

 

[alexsok]

Rumours say we should look forward to the official annoucement tommorow - everything should clear up then...

 

[MfA]

Well s3graphics.com seems offline, so might have been hacked.

 

[Humus]

It's up, just slow.

 

[OpenGL guy]

Advanced Deferred Rendering
Hierachical Z
In addition to conventional z buffering, where z values are computed and compared at the
pixel level, Columbia supports another coarser z buffering scheme in advance of the
standard z buffering stage. An incorporated zero-cycle clear and optional deferred
rendering (two pass) mode further optimizes color/z-buffer memory traffic and
pixelperformance. Deferred rendering virtually eliminates opaque (redundant) back-to-front
overdraw and saves also precious conventional z write traffic.

"2 passes"... sounds like forcing things to render "Doom 3"-style, i.e. lay down Z data first, then do everything else.

I don't see much novelty here...

 

[T2k]

It's up, just slow.

Hope only the website is slow...

 

[Mr.huang]

http://www.s3graphics.com/index.html

"Alpha Chrome" is Discrete Moblie Chip?

 

[Dreamer]

http://www.s3graphics.com/index.html

"Alpha Chrome" is Discrete Moblie Chip?

Alpha Chrome = SavageXP

Alpha Chrome = Mobile
Savage XP = Desktop

 

[Psikotiko]

Anyone knows the specs?

The only thing that s3 says is "Coming Soon to a System Near You...
"