View Full Version : NV35 Still 4x2?
surfhurleydude
04-30-03, 09:47 AM
I think I see Hellbinder showing fear over the NV35 now... He's scared. What the hell is he going to bash now?
I know already. He's going to bash the fact that even though the NV35 is 25% faster than the Radeon 9800 Pro in AA situations, it doesn't look as good. Pretty see through IMO.
Hey but Radeon AF is faster than nVidias but doesn't look as good either.
The hypocracy continues.
digitalwanderer
04-30-03, 10:28 AM
Originally posted by surfhurleydude
I think I see Hellbinder showing fear over the NV35 now... He's scared. What the hell is he going to bash now?
I don't think Hellbinder fears any video cards, he just says what he thinks/knows about 'em. :rolleyes:
As for what is he gonna bash? Hell, if nVidia doesn't give him ample ammo I'll give him some me own self! :p
Originally posted by Hellbinder
Nv35 is going to ship at 400 Core and either 350, or 400ish DDR.
Both Core and Memory are clocked higher than that on the final retail cards.
Richthofen
04-30-03, 01:57 PM
i think it will be 450/450 for the ultra like inquire said.
Who really cares if it's 4x2 or 8x1? DX9 doesn't need 8x1 to work right so who cares so long as it offers decent performance :)
Based on the current info, it will offer more than a decent performance.
Originally posted by Dazz
Who really cares if it's 4x2 or 8x1? DX9 doesn't need 8x1 to work right so who cares so long as it offers decent performance :) It all depends on if/when cards move to fewer texture layers. An 8x1 config will perform much like a 4x2 in current games. But in a game with a single texture layer, an 8x1 card would perform twice as fast as a 4x2 because the 2nd texture unit would sit idle.
Well, I must admit, I think the whole idea of "8x1", "4x2", ... is kinda too old now.
Right now, you'd be better of by saying, for example:
R300: 8T+8P ( Texture + FP )
NV30: 8T/4P+8X ( Texture/FP + FX )
NV25: 8T+4X
R200: 8T+4X
That kind of stuff. It's a lot more accurate IMO.
Also, you may want to add some other things, like precision. So, it could be:
R300: 8T+8P24
NV30: 8T/4P32+8X12
NV25: 8T+4X9
R200: 8T+4X16
Of course, then, everyone who isn't technical would get lost nearly immediatly. But hey, that isn't MY fault! :)
Uttar
You lose me every time you type ;)
surfhurleydude
05-03-03, 03:29 PM
Yeah, Uttar, I'm lost.
I still want to know how much better the NV35 is than the GeForce FX and if it's single texturing is a lot better than it was with the NV30 though. :o
bkswaney
05-03-03, 11:17 PM
I think the NV 35 will rock. The NV 30 will get a nice speed boost by the Det FX drivers in AA.
NV says they have somekind of debug code running in the drivers we have now and the new drivers will get a huge speed jump in AA. The debug software is being removed on the 50+ drivers. :D
Also from some of the newer reviews coming out the IQ is really coming up with the newer drivers.
The NV35 will rock! :beer: :jumping:
GlowStick
05-04-03, 12:43 AM
Originally posted by bkswaney
I think the NV 35 will rock. The NV 30 will get a nice speed boost by the Det FX drivers in AA.
NV says they have somekind of debug code running in the drivers we have now and the new drivers will get a huge speed jump in AA. The debug software is being removed on the 50+ drivers. :D
Also from some of the newer reviews coming out the IQ is really coming up with the newer drivers.
The NV35 will rock! :beer: :jumping:
Awesome
Demirug
05-04-03, 05:31 AM
Originally posted by Uttar
Well, I must admit, I think the whole idea of "8x1", "4x2", ... is kinda too old now.
Right now, you'd be better of by saying, for example:
R300: 8T+8P ( Texture + FP )
NV30: 8T/4P+8X ( Texture/FP + FX )
NV25: 8T+4X
R200: 8T+4X
That kind of stuff. It's a lot more accurate IMO.
Also, you may want to add some other things, like precision. So, it could be:
R300: 8T+8P24
NV30: 8T/4P32+8X12
NV25: 8T+4X9
R200: 8T+4X16
Of course, then, everyone who isn't technical would get lost nearly immediatly. But hey, that isn't MY fault! :)
Uttar
I am not sure about R200 but NV25 use 8T+8X9. There are 2 Register Combinder for each pipeline.
Originally posted by Demirug
I am not sure about R200 but NV25 use 8T+8X9. There are 2 Register Combinder for each pipeline.
Yeah, realized that, thanks for noting it :)
Sorry for the mistake. In fact, I later posted something similar to B3D, and there I didn't do the error :)
Uttar
StealthHawk
05-04-03, 06:40 AM
Originally posted by bkswaney
I think the NV 35 will rock. The NV 30 will get a nice speed boost by the Det FX drivers in AA.
When did nvidia say this?
I'm still a firm believer that the NV30 has fine 2x FSAA. It has a very low performance hit.
I'm also a firm believer that all nvidia cards to date starting from the gf3 have poor 4x FSAA quality. Not worth the perfomance hit at all. Maybe if nvidia could get the hit down below 20% you could justify using it.
As a side question, is nvidia still using the OG modes for 6x and 8x in the driver control panel, or are they using the "true" RG modes which look a hell of a lot better?
It all depends on if/when cards move to fewer texture layers. An 8x1 config will perform much like a 4x2 in current games. But in a game with a single texture layer, an 8x1 card would perform twice as fast as a 4x2 because the 2nd texture unit would sit idle.
Where does this information that future games are gonna be single-textured come from? Its one of the most rediculous things I've heard in ages.
Future games will use more textures, not less. Why do you think the specs from gf4 to gf-fx/9700 was from 4 TPP* to 16 TPP*
And under dual texturing situations, a 4x2 architecture offers equal performance to a 8x1 architecture. (Well roughtly, give or take lots of various factors)
Having 8x1 when there are no textures, i.e. stencil volumes, Z writes (Alot of engines these days do a specific Z write first, before anything else), and then 4x2 when the heavy texturing comes in is a great innovation. I dont know why ppl are soo keen to bash it.
* Textures Per Pass.
Originally posted by Nutty
Having 8x1 when there are no textures, i.e. stencil volumes, Z writes (Alot of engines these days do a specific Z write first, before anything else), and then 4x2 when the heavy texturing comes in is a great innovation. I dont know why ppl are soo keen to bash it.
Well, because of having only 4 pipelines, the NV30 got execrable FP performance. I think that's already one good reason to bash it. Although yes, people probably bashed it more than necessary...
Uttar
I thought it outperformed 9700 with recent drivers?
Regardless of whether they "cheat" or not, they are WHQL and the bright sky is back correctly in 3dmark03.
Like you've told me before Uttar, there are many things that went wrong with nv30. I dont think the design of the core is one of these things tho.
Given nv35 will use the same system, just with better bandwidth/things fixed, we'll see how the adaptive approach should work in a correctly working card when nv35 gets reviewed.
I'm praying they dont mess this one up. :D
Originally posted by Nutty
I thought it outperformed 9700 with recent drivers?
Regardless of whether they "cheat" or not, they are WHQL and the bright sky is back correctly in 3dmark03.
Of course, but the point is that they're using FX12. While they may be able to use FX12 in a lot of places and get great performance thus, it isn't possible to do so everywhere. And in cases where it isn't, the NV30 is 2 times slower clock for clock than the R300/R350.
Uttar
I actually very much doubt there are using INT12 for operations that probably require fp16 minimum. Hell even gf4 used to use 11bits of precision internally IIRC. So its only an extra bit on top of that. When in reality the app probably requires much more. Hence I think fp16 is probably the minimum being used.
I know for a fact that all texture addressing is done in full 32bit mode, so its not everything that is dropped down to lower precision.
But at the end of the day, it shouldn't matter what the hardware does behind the scenes. As long as it looks good and correct and is fast, it really just doesn't matter. And it shouldn't either.
Supposedly from what I've heard multiple times, texturing as we know is going the way of the dodo, and the pixel shaders will start to be used to create texturing.
Not for a long time. Why waste hundreds of clock cycles calculating a value, when you can just look it up in a texture?
Even then, how the hell do you convert your artists amazing textures into procedural algorithms? You gonna write a procedural routine for every texture in the game?! No way man. It aint gonna happen.
The textures that will be replaced will be things like normalization cube maps, gradient bars, BRDF tables, Anisotropic lighting functions. These are the kind of textures that will be replaced with calculations on the fly.
But as more and more complex lighting systems start being used, then more and more textures are gonna be used, to support them. You think sub surface scattering settings will use a procedural routine written by some coder, or use the settings done by an artist and stored in a texture? Answer: It will be the artists texture.
At the moment its still faster to use a normalization cube map, that it is to normalize in a fragment program. And quite probably textures will remain the faster of the two routes for complex lighting equations.
Artist created assets will remain as textures though. The hardware applying them will go through many transformations, but they will still be textures.
And there will probably never be a day when there is only 1 single texture applied to stuff. Even with the function lookup textures replaced.
vBulletin® v3.7.1, Copyright ©2000-2012, Jelsoft Enterprises Ltd.