Nv40 16 full pipelines- The Inq.

GF256 does not support quad texturing...
GF3 was the first nV part that can do that.
They are probably confused it with the Savage 2000...

 

Sorry to go against the flow

I don't think NV are above such mistakes. Am I the only one that believes NV30 was suposed to be a 4X2/8X1?

GF1 DDR wasn't under any circumstances bandwidth starved.

It was pretty heated at the launch Savage 2000, Matrox G400, dual rage 128...

 

Yes, you are.

No, actually there some people - those people don't realize the significance of the GPU operating on quads.
R300 has two parallel pixel engines both can operate on one quad / cycle.
NV30 has one pixel engine that can operate on one quad / cycle.
It's not really possible to do 8 pixels/clock this way.

Z units are totally different and are usually operate faster anyway (since R100 and NV20...)
The only new thing in NV30 is that it can do Z update operations 8/cycle when the pixel engine is completely disabled.

 

Seems that Nv will be talking of PS/VS 3.0 at the Game Developer Conference 2004 and one of the presentation will be Cinematic Effects II: The Revenge
http://www.hardware.fr/news/lire/05-03-2004/

 

I'm still waiting for Cinematic Effects I: Where the hell are they?

 

Well it died with the NV30, didn't it?

 

We already have cinematic effects in my opinion the problem is that the pixar's of the world don't sit on their laurels and make films with the same technology of the 80s if they did people would be aghast b/c they would think gee whiz my computer game looks this good ...

 

Who says it's one quad?

Yes it is. Not easy, not likely, not true, not happening... but it is possible.

So Z units aren't on the pixel engine? Don't they operate after quads are dispatched? because if they do that means they operate on two quads at once. And I think they are more per pixel pipeline, not faster.

I thought Z units operate on pixels, not geometry. pixel engine can't be disabled when it is operating on pixels. maybe just the shading part. I think that's a bug. I don't have any proof but I didn't pretend I was right.

Ps please excuse my English, and keep in mind I'm not flaming

 

No.

Take the NV20 for an example.
It has 16 early Z units.
It takes an area of 4x4 pixels (when not multisampling) and decides which is visible, which isn't.
That covers 4 quads. Only after this visibility check passed for a quad does the quad gets dispached to "pixel processing" (TMUs + register-combiners).

Z units operates on 4x4 subpixels when multisampling it's just covers less pixels (8 pixels = 2 quads with 2xMSAA, 4 pixels = 1 quad with 4xMSAA).

NV20 can however only update the Z value of 4 pixels even when there's no multisampling.
NV30 can update the Z value of 8 pixels (IIRC it can only do that with no MSAA or 2xMSAA, not 4xMSAA).

 

I seriously doubt you need 16 early Z units for the hierarchical Z. You only care about the largest/smallest Z value for the group of pixels you are looking at...

Because there are only 4 Z units.

My understanding was that it could update 16 Z values with AA enabled.

 

Heh, that's great

 

Um, when were we talking about Hierarchical Z? The post you responded to was describing the NV20.

 

Yes, NV20 has 16 Z-compare units that can discard at most 16 pixels per clock. Of course if you can't discard, the pixel pipeline will be the limiting factor.

 

I recently asked Kirk about this. The early rejection scheme they emply, according to him, is onchip and bears similarities to Hier-Z but is organised differently - if this were the case the system would probably sit between the setup and pixel engines and I'm not sure how much it has to do with the number of z-ouputs.

 

I thought NV20 had hierarchical Z obviously.

 

Well, It doesn't have hierarchical Z it's an ATI only thing.
ATI cards are generally faster at pure front-to-back rendering because IIRC HZ can discard somewhere in the 64 pixels/clock range.
Of course updating the hierarchical Z has it's own cost.

NV has the 16 Z compare units. Z-compression operates on 4x4 tiles, so if the tile is compressible the bandwidth is sufficient to feed the Z units even on a NV20.

 

I didn't think nVidia implemented z-compression until the NV30.

 

IIRC Z-compression was with NV25, colour compression came with NV3x

 

With regards to the NV20 and z-compression; from the NVIDIA website:

http://www.nvidia.com/object/techbrief_lmaii.html

 

Ah, thanks.