Any more info on the nv40?

[Geforce4ti4200]

There should be preliminary specs by now. Will it be an overclocked fx5900u or a totally new beast? Will any be available before 2004? will it be .13 microns or the old .15? will it use 8x2, 16x1 or even 16x2? will the core be clocked past 500MHz, say 600MHz for example? what about the ram? 1100? 1200?

600/1200, 8x2 and early 2004 is my guess and $550 for a 256mb

[Nebuchadnezzar]

32x1 Pipes
Pixel & Vertex Shader 4.0+++
0.09µ
1Ghz Core
32MB Embedded RAM
512bit DDR3 bus
LMA4
1600Mhz (800) DDR3 RAM
456M Trannies
And lot more things but i can't say anymore now!












....:rolleyes:








:p

[Paul]

- Totally new (or as new as any core is over the previous generation).
- Available in December 2003, in limited quantities at best. Mass availability next year.
- .13 Microns
- 8x1 is the most likely bet, 8x2 is possible.
- Core at 600, Memory at 1200-1400mhz (600-700mhz).
- GDDR2, probably.
- Same price as any new core... expensive ;)

That is mostly speculation. The core/mem speeds were from Uttar, and it was posted on the front page of this site at some point.

[Geforce4ti4200]

600/1200 was my guess. From what it sounds, the nv40 is nothing more than a higher clocked nv35 which was 450/850(or 900) so were talking a third faster to be exact on both core and ram. Perhaps Nvidia is saving the cherries that make 600MHz for the nv40 and use the same ram that was on the fx5800u

[BigFish7]

:mad: WAYT ARE YOU TALKING ABOUT?!?!?!
the nv40 should be totaly different from what we have ever seen befor.
(although the real change should come with the NV45 or more likely NV50).
don't even think about an "OCed core" :mad:
i'm allmost 100% sure that nVidia will show new AA patterns with new AA technology.

[Nebuchadnezzar]

Originally posted by BigFish7
:mad: WAYT ARE YOU TALKING ABOUT?!?!?!
the nv40 should be totaly different from what we have ever seen befor.
(although the real change should come with the NV45 or more likely NV50).
don't even think about an "OCed core" :mad:
i'm allmost 100% sure that nVidia will show new AA patterns with new AA technology.

Slown down there.. no need to get agressive.

[Nutty]

As long as its got ;

a) Good PS2 performance. Really good in FP32
b) Improved IQ in AA. (Muchly needed)
c) Doesn't sound like a jet fighter.

I reckon it'll be a good'un.

[BigFish7]

Originally posted by Nutty
As long as its got ;

a) Good PS2 performance. Really good in FP32
b) Improved IQ in AA. (Muchly needed)
c) Doesn't sound like a jet fighter.

I reckon it'll be a good'un.

in addition to its DX9.0 killer performance :D

[gmontem]

The NV40 is gonna outperform the NV35 by at least two fold in the new PRMark benchmark.

[Uttar]

I'd actually love to know whether FP16 is gone in the NV40. Because that'd also mean the register usage problem is gone.

Some people say they think it's gone, but they aren't too sure themselves ( this isn't speculation though, it is rumors ) - frankly, I don't know. I hope it's gone, and that's all I can say.


Uttar

[GlowStick]

I really cant see nvidia going 8x1, they have stuck to their 4x2 for so long i just dont see why they would change, but i guess they could if they wanted to!

just my 2 cents!

[Geforce4ti4200]

If nvidia uses 4x2 or 8x1 they will suck because the radeon 9700 and up uses 8x1 and I wouldnt be supprised if Loci or r420 uses 8x2

[SurfMonkey]

Originally posted by Uttar
I'd actually love to know whether FP16 is gone in the NV40. Because that'd also mean the register usage problem is gone.

Some people say they think it's gone, but they aren't too sure themselves ( this isn't speculation though, it is rumors ) - frankly, I don't know. I hope it's gone, and that's all I can say.


Uttar

I've heard it's gone. But I'm also hearing some funny, and contradictory, things as well. Still it sounds as though the NV3x debacle hasn't really affected the NV4x.

[Uttar]

Originally posted by SurfMonkey
I've heard it's gone. But I'm also hearing some funny, and contradictory, things as well. Still it sounds as though the NV3x debacle hasn't really affected the NV4x.

I've also heard that, but I'm really not too sure how credible it is.

What I'm more interested in, however, ( partly because the odds of getting those details in the next few weeks/months is higher ) , is the microarchitecture of the NV4x's pixel pipelines. That is, number of FPUs, TMUs, ...
Then, knowing what those FPUs can do becomes more interesting, because a part which can do COS/SIN in one cycle but only got 1 FPU isn't as interesting as one who does it in four cycles but got 8 FPUs :)


Uttar

[Nutty]

I think FP16 will still be there. Nv has spent too much money and time getting ppl to use. Doom3 uses it, therefore if NV40 can do FP16 faster than FP32, their doom3 benchmarks, (and all future games that use the doom3 engine) will be higher on that.

It would be nice to see 8x2, with the ability to do twice the number of FP16 instructions than FP32 again. But with FP32 performance amazingly good in the 1st place, so FP16 is even better.

You dont need full 32bit per channel precision for everything. Not for a good few years.

Also given NV3X is 4x2, but does 8x1, in Z and stencil. It would be amazingly cool if NV40 was 8x2, and did 16x1 in Z and stencil. Think of the huge stencil speed if it did 16 stencil fragments in parallel. This would massively help engines like doom3 that use alot of stencil volume shadows. The fillrate speed would be phenominal.

[Geforce4ti4200]

so a 16x1 is superior to an 8x2? Maybe Nvidia better supprise us before the 16x2 r420 or "loci" does ;)

[Paul]

I'd be very suprised if the NV40 maintained any kind of 4x2 at the expense of full 8x1. I'd be even MORE suprised if Loci had 16x2.

Hell, i'll say it now. It's not going to happen :P

[Nutty]

so a 16x1 is superior to an 8x2?

Yes, in single texturing or not texturing, and operations that dont require textures. Stencil and Z writing for instance.

[Uttar]

But a 16x1 is also more expensive to implement, transistor-wise, than a 8x2.
But personally, I don't like that sort of generalization at all. It's outdated IMO, because a 8x1 can be tremendously more powerful than another 8x1 at identical clock speeds, for example if it got two FPU units instead of one - then it's a lot more powerful in shaders.


Uttar

[Nutty]

What I'm saying is, nv claim their 4x2 operates as 8x1 in Z and stencil. Now if they go ahead and create an 8x2 architecture, would it be possible for them to make it operate like a 16x1 in Z and stencil also.

Having 16 fragments processed in parallel in Stencil is gonna make toast of doom3 based benchmarks. And we all know benchmarks means prizes.. :)

[Geforce4ti4200]

16x2 is gonna happen oneday for sure, but no one knows when :( also isnt it almost as bad to keep ramping up raw MHz? also isnt memory bandwith more important than anything?


edit:

ti4200 is 4x2, fx5600u is 4x1 so its slower in many benchmarks but because of its 800MHz ram, it wins in fsaa and high res due to bandwith.

[Nebuchadnezzar]

[i]ti4200 is 4x2, fx5600u is 4x1 so its slower in many benchmarks but because of its 800MHz ram, it wins in fsaa and high res due to bandwith. [/B]

:rolleyes: :rolleyes: :rolleyes: :rolleyes: :rolleyes:

[nvnews-reader]

Originally posted by Uttar
But a 16x1 is also more expensive to implement, transistor-wise, than a 8x2.
But personally, I don't like that sort of generalization at all. It's outdated IMO, because a 8x1 can be tremendously more powerful than another 8x1 at identical clock speeds, for example if it got two FPU units instead of one - then it's a lot more powerful in shaders.


Uttar

How much more performance do you think ati and nvidia can get without moving to a wider pipeline?

[StealthHawk]

Originally posted by Geforce4ti4200
also isnt memory bandwith more important than anything?

No, how do you figure that? What I mean is, you need adequate bandwidth or else your card will be slower than its potential, but adding excess bandwidth won't help you either. It's a balancing game. Saying that bandwidth is the end all be all of performance is incorrect.

ti4200 is 4x2, fx5600u is 4x1 so its slower in many benchmarks but because of its 800MHz ram, it wins in fsaa and high res due to bandwith.

It wins in FSAA because it has color compression.

[Geforce4ti4200]

true, but without fsaa and high res the ti4200 wins badly :)