GDC2004 Nvida Presentations

[pocketmoon_]

What can we gleen from these ?

http://developer.nvidia.com/object/gdc_2004_presentations.html

For starters, in OpenGL land, NV4x can do 4 texture lookups in a vertex shader.

NV40 is MIMD
So branching is fast I would have thought that mean quite a few extra transistors! Jumping from SIMD, where you have duplicate sets of registers but run the same instructions against them.

 

[991060]

I'm wondering why ATi is so low-pitched this time, only 2 presentations so far.

 

[Demirug]

What can we gleen from these ?

http://developer.nvidia.com/object/gdc_2004_presentations.html

For starters, in OpenGL land, NV4x can do 4 texture lookups in a vertex shader.

NV40 is MIMD
So branching is fast I would have thought that mean quite a few extra transistors! Jumping from SIMD, where you have duplicate sets of registers but run the same instructions against them.

You can use 4 different textures and make up to 512 lookups.

The Vertexunit ist MIMD. The Pixelunit(s) still use SIMD.

 

[pocketmoon_]

Thanks for claifying that.

I wonder if each vertex unit is fully MIMD or if the number of available texture units decreases with each code branch.

 

[Demirug]

Thanks for claifying that.

I wonder if each vertex unit is fully MIMD or if the number of available texture units decreases with each code branch.

You can use allways up to 4 textures. There is no dependency with code branches.

 

[pcchen]

It seems floating point filtering (trilinear/anisotropic filtering, including cube map) and blending are supported in NV40, although maybe limited to FP16. That's pretty good IMO.

 

[ben6]

Assuming the presentations are accurate this looks like a nice update feature-wise if performance can be up to par.

 

[991060]

Hmm, is there any interesting info about next-gen consoles? I'm really looking forward to.

 

[mreman4k]

I'm wondering why ATi is so low-pitched this time, only 2 presentations so far.

I do not think ATI has to prove themselves, they have had a generally successful 18 months. NVidia on the other hand has the NV30 situation in its immediate past, so they've got to come back with something.

 

[John Reynolds]

The AA pdf was a little disheartening because it looks like the only real way to consistently AA a DX9-class game is to use super-sampling.

 

[DemoCoder]

Other info gleened from presentations:

NV40 supports Multiple Render Targets (yes!)
NV40 has hardware normalize instruction (NRM)

Multi-precision still a part of the architecture.

 

[Bjorn]

One thing that i noticed is that NVidia clearly doesn't shy away from talking about Ati extensions. They're all over the place in the PDF's.

 

[Joe DeFuria]

One thing that i noticed is that NVidia clearly doesn't shy away from talking about Ati extensions. They're all over the place in the PDF's.

I noticed that too. Interesting what happens when someone doesn't actually charge you for using their extensions...

 

[Bjorn]

One thing that i noticed is that NVidia clearly doesn't shy away from talking about Ati extensions. They're all over the place in the PDF's.

I noticed that too. Interesting what happens when someone doesn't actually charge you for using their extensions...

That might have something to do with it

But it's at least good to see that they aren't stubborn about creating their own extensions for everything.

 

[Ante P]

The AA pdf was a little disheartening because it looks like the only real way to consistently AA a DX9-class game is to use super-sampling.

At first when I started reading it I thought there'd be a point with some swift cool new sollution at the end, but no.

Breaking AA or resorting to SSAA is not the way to go in 2004. I'd say screw fancy DX9 stuff if I can't have my AA.

 

[oddfellow]

The AA pdf was a little disheartening because it looks like the only real way to consistently AA a DX9-class game is to use super-sampling.

At first when I started reading it I thought there'd be a point with some swift cool new sollution at the end, but no.

Breaking AA or resorting to SSAA is not the way to go in 2004. I'd say screw fancy DX9 stuff if I can't have my AA.

I'm with you on that

 

[Johnny Rotten]

The AA pdf was a little disheartening because it looks like the only real way to consistently AA a DX9-class game is to use super-sampling.

Poor IHV's. Chips feature Supersampling and the masses cry 'Too slow! Too slow!". IHV's move to multisampling and the masses cry "We want the whole frame antialiased!", and now with the apparent move back to supersampling the complaints are beginning anew. Poor buggers cant win.

 

[StealthHawk]

The AA pdf was a little disheartening because it looks like the only real way to consistently AA a DX9-class game is to use super-sampling.

Poor IHV's. Chips feature Supersampling and the masses cry 'Too slow! Too slow!". IHV's move to multisampling and the masses cry "We want the whole frame antialiased!", and now with the apparent move back to supersampling the complaints are beginning anew. Poor buggers cant win.

It's too bad that Supersampling is too slow!

 

[ben6]

Yah know what? I want good looking antialiasing period! Yah know what impressed me first about modern computer hardware? Watching a Voodoo 5 5500 run right next to the exactly same equipped system with a Geforce2 GTS in Need For Speed High Stakes

 

[nutball]

It seems floating point filtering (trilinear/anisotropic filtering, including cube map) and blending are supported in NV40, although maybe limited to FP16. That's pretty good IMO.

I'm wondering why (and if) these are limited to FP16 :?

I mean, in for a penny, in for a pound, as they say here in Blighty.

Also, how are they implemented? Blending a la the FF pipeline? Frame-buffer readback? Current frame-buffer contents available in PS?

So many questions, so little clue