nv40 excitement growing ?

[digitalwanderer]

Originally posted by The Baron
The 8x2 thing is pretty much confirmed.
Uhm, "pretty much"?!? :lol:

C'mon Baron, it either IS confirmed or it IS NOT confirmed..."pretty much" is a pretty weanie way of saying "mebbe". :rolleyes:

How comfortable would you feel with your girlfriend telling you she's pretty much not pregnant? :eek2:

and dig, it's like I said--there are NV40s in the wild.
Suuuuuuure there are Baron, suuuuure there are.

And it's a good thing too, because I'm still ready. :p

[The Baron]

Dig, you know we're reaching a point where pipeline configuration is meaningless? That point is R500/NV50. NV40/R420 pipeline configuration is almost meaningless. That's why it's "pretty much" confirmed--there's no hard and fast way (that I know of) to tell from the data available.

[Paul]

Originally posted by digitalwanderer
How comfortable would you feel with your girlfriend telling you she's pretty much not pregnant? :eek2:

http://files.lasalas.net/pauls/nvn/coathanger.gif

Suuuuuuure there are Baron, suuuuure there are.

I can understand your scepticism after the NV30, but I assure you samples are out there. A couple of high profile european devs have had them, for example.

[retsam]

you friggin guys are killing me ..... dont we deserve numbers ....... what are there reactions to the nv40 positive or negative ...do tell dooooo telll!!!





rets:afro2:

[digitalwanderer]

Originally posted by The Baron
Dig, you know we're reaching a point where pipeline configuration is meaningless? That point is R500/NV50. NV40/R420 pipeline configuration is almost meaningless. That's why it's "pretty much" confirmed--there's no hard and fast way (that I know of) to tell from the data available.
Then why are y'all so adament that it's an 8x2? On what are you basing that?

[Mariner]

Well, the only question now is do we know that NV40 is definitely 8x2/16x1 or is that just what NV told the devs! :D

When the NV40 is released, are we going to have to run fillrate testers and the like to discover the actual pipeline layout or can we believe what the press releases say this time?

:afro:

[Paul]

Originally posted by Mariner
When the NV40 is released, are we going to have to run fillrate testers and the like to discover the actual pipeline layout or can we believe what the press releases say this time?

For the most part, the Internet GPU community has changed since the NV30 and the months of nVidia debacles. Rest assured that people will be going over the cards (both nV and ATi) and drivers with a fine toothcomb. Whilst some sites might continue the old practise of just reeling out the PR literature supplied by the IHVs, the likes of B3D won't.

[aapo]

Originally posted by The Baron
it's 8x2/16x1.

That's not even possible, there's some kind of confusion now. It's either 8x2/16x0 or 16x1 or 6x2/12x0 or whatever, but not 8x2/16x1. It wouldn't make any sense whatsoever to run a chip capable of 16x1 in 8x2 mode - there's nothing to gain with 8x2 configuration compared to 16x1 configuration.

[Uttar]

The NV40 is 6x2/12x1/8x2/16x1. Marketed as a 16 pipelines design. I did not say 12x0 or 16x0; although maybe it isn't capable of 12x1/16x1 and just 12x0/16x0. I doubt that though.
The NV41 is 3x2/6x1. Marketed as a 6 pipelines design.

Aapo: There is. See the NV31, NV34, and NV36 which are all 2x2/4x1. They run in 2x2 mode most of the time, but they can run in 4x1 mode in very specific cases. That's because it's still less expensive to implement a "specific case" 4x1 mode, I assume. Furthermore, both the NV3x and NV4x are Xx2 architectures, so there would be some serious wastage going in in 4x1/16x1 mode if it was used everywhere.

And I could be wrong on this, of course, because this info is such a fricking mess :s


Uttar

EDIT: If this information is correct, I also believe all *current* 100% synthetic tests (I'm not talking about 3DMark here; more like ShaderMark) are not going to tell the whole story anymore. But remember I'm not guaranteeing anything anymore :p

[digitalwanderer]

Originally posted by Uttar
The NV40 is 6x2/12x1/8x2/16x1. Marketed as a 16 pipelines design. I did not say 12x0 or 16x0; although maybe it isn't capable of 12x1/16x1 and just 12x0/16x0. I doubt that though.
The NV41 is 3x2/6x1. Marketed as a 6 pipelines design.

Aapo: There is. See the NV31, NV34, and NV36 which are all 2x2/4x1. They run in 2x2 mode most of the time, but they can run in 4x1 mode in very specific cases. That's because it's still less expensive to implement a "specific case" 4x1 mode, I assume. Furthermore, both the NV3x and NV4x are Xx2 architectures, so there would be some serious wastage going in in 4x1/16x1 mode if it was used everywhere.

And I could be wrong on this, of course, because this info is such a fricking mess :s


Uttar

EDIT: If this information is correct, I also believe all *current* 100% synthetic tests (I'm not talking about 3DMark here; more like ShaderMark) are not going to tell the whole story anymore. But remember I'm not guaranteeing anything anymore :p
Dude, you don't have to guarantee anything as that info just confuses the bejesus out of me more than anything else! :confused:

[Demirug]

Originally posted by digitalwanderer
Dude, you don't have to guarantee anything as that info just confuses the bejesus out of me more than anything else! :confused:

Think about it in this way:

Let us say that NV40 has 4 processors. 3 pixelprocessors and 1 vertexprocessor. Each processor work on up to 4 units at the same pipeline stage. In this case they maximum output is 4 units per clock. A unit can be a pixel or a vertex. Let's say that the vertexprocessor can work on vertex and pixel units. A pixelprocessor works only on pixels. If this is right NV40 can work on up to 16 Pixel (or 4 quads). In the past this means that you have a 16 pipeline chip. But this does not mean that you can output 16 pixel per clock. This will depend on the number of ROPs. On the other hand the normal speed of a pixelprocessor would only 2 units because each pixelprocessor works in defaultmode as 2x2 (2 pixel with 2 Texops + some other ops). Only in special case it can do 4 pixel (nvidia call this doppel pumped).

As Uttar I have to commit that this all can absolute wrong but IMHO it sounds good.

[picofarad]

So if all of this is right, is this related to texture samplers in vertex shader3.0? When you want to vertex texture you loose resources on the pixel side (8x2 -> 6x2)?

[Demirug]

Originally posted by picofarad
So if all of this is right, is this related to texture samplers in vertex shader3.0? When you want to vertex texture you loose resources on the pixel side (8x2 -> 6x2)?

No, you have to think in the other way. Normaly it is 6x2 but if they vertexprocessor have nothing to do it will calculate some pixel quads. During this time we have a 8x2 chip. This is a simplification because in realitiy it will be more complex.

[The Baron]

EDIT: If this information is correct, I also believe all *current* 100% synthetic tests (I'm not talking about 3DMark here; more like ShaderMark) are not going to tell the whole story anymore. But remember I'm not guaranteeing anything anymore
then okay, how long before somebody builds a better synthetic benchmark? :p

[Uttar]

One of the things I wonder is just how identical the VS and PS "processors" are. I'd assume they're 100% identical, but that'd seem strange, as the focus shouldn't be on the same things (texturing in the VS doesn't merit to be as fast as in the PS; branching in the PS doesn't merit to be as fast as in the VS; and I'd also say, but that's just IMO, that ADD is more important in the PS than in the VS).
If it isn't identical, I'd assume the differences would solely be in the number of units; not in the type of them (one unit wouldn't be "less efficient" - there'd simply be less of them) in order to reduce the part's R&D time.

It would also seem logical to me that each "processor" can work on a different program. Being able to switch to the next program before all 12 pixels have been rendered seems like a slight, but interesting, improvement to me.
Once again, if it's all wrong, I'm not responsible for this :p

What I'm 99% sure of, however, is that there IS a 16x1 mode for the NV40, and there IS a 6x1 mode for the NV41. The rest is all about melting contradictory information, obviously.


Uttar

[nobie]

So, NV41 is basically 4x2 that can sometimes be 6x2?

[DMA]

Originally posted by digitalwanderer
Dude, you don't have to guarantee anything as that info just confuses the bejesus out of me more than anything else! :confused:

I know what you mean. Reading all this technical stuff makes my brain hurt. Now i know how Homer feels. :D
http://www.photocountry.com/users/14308/homersbrain.jpg

[Uttar]

Originally posted by nobie
So, NV41 is basically 4x2 that can sometimes be 6x2?

No, it's fundamentally 6x2. Sometimes it's 12x1. Then, in other cases, it can be 8x2. And if BOTH the 12x1 and 8x2 conditions are met, it's 16x1.


Uttar

[-=DVS=-]

Originally posted by Uttar
No, it's fundamentally 6x2. Sometimes it's 12x1. Then, in other cases, it can be 8x2. And if BOTH the 12x1 and 8x2 conditions are met, it's 16x1.


Uttar

To put it simple NV40 pipeline range is from 6x2 to 16x1 :rolleyes:

I guess R420 woun't be as complex chip then for sure :confused:

[digitalwanderer]

Originally posted by -=DVS=-
To put it simple NV40 pipeline range is from 6x2 to 16x1 :rolleyes:

I guess R420 woun't be as complex chip then for sure :confused:
Well you know that ATi always does the "brute force" approach and nVidia prefers the "elegance of flexibility".... :rolleyes: :lol:

[MUYA]

Originally posted by Uttar
No, it's fundamentally 6x2. Sometimes it's 12x1. Then, in other cases, it can be 8x2. And if BOTH the 12x1 and 8x2 conditions are met, it's 16x1.


Uttar

Will this complexity just hurt its performance or will NV actually be able get things in order? or..something....16X1? 8X2? 6X2? man....heads gonan start hurting soon

[Paul]

Originally posted by digitalwanderer
Well you know that ATi always does the "brute force" approach and nVidia prefers the "elegance of flexibility".... :rolleyes: :lol:

Well, you roll your eyes but I think there is something to be said for the approach nVidia tried to take, with allowing developers to choose the appropriate precision for the task - as you know, there are many areas when FP16 (or even less) is plenty. The problem is only that it was implemented badly, and that FP16 wasn't sufficiently faster than FP24. Coupled with the fact that nVidia needed to force lower precision on to get decent framerates, it's given the whole idea a bad name which is completely unjustified.

[The Baron]

Originally posted by MUYA
Will this complexity just hurt its performance or will NV actually be able get things in order? or..something....16X1? 8X2? 6X2? man....heads gonan start hurting soon
no, it won't hurt performance... or shouldn't. maybe a bit, but I don't think this is another NV30 fiasco. the Crazy Pipelines stuff has been on its way for quite a while now--Series 5 will have similarly goofy pipelines (at least in theory), R500 will definitely not be accurately described by its pipelines along, and NV50 will probably be the same.

[MUYA]

Originally posted by The Baron
no, it won't hurt performance... or shouldn't

well guess we will find out next month soon enough on games and synthetics.....its the waiting....


erm yes

/me goes to cook some Noodles

[digitalwanderer]

Originally posted by MUYA
/me goes to cook some Noodles
Mmmmmmmmmm...............................noodles.. ......................mmmmmmmmmmmmmm.