NV40 3DMark 2003 scores revealed -theinquirer

[MMJ]

http://www.theinquirer.net/?article=15169

Looking good (if true).

 

[nobie]

That's over twice as fast as a GFFX 5950!

 

[tEd]

i'm waiting for a certified driver score if you don't mind

 

[retsam]

i think i need to go and shread my credit cards..like right now

 

[Ailuros]

Perhaps we should all now be expecting briefings by Nvidia's competitors that 3DMark03 isn't now a valid indicator of 3D graphics performance?

Good point or a vice versa version of that theory.

Considering that 7k points are reachable with today's ultra high end systems I wouldn't say that 10 or 12+K scores are unexpected.

 

[Dave Baumann]

Remember the composition of GT2 & 3.

 

[vb]

PS1.4!!???

OMG. What about 3DMark04/3DMark03SE ?

 

[tEd]

PS1.4!!???

i think he means the stencil stuff

 

[nobie]

The thing to remember with 3dmark is that it's not just a video card test, it stresses alot of different component and is often CPU limited. I wouldn't be surprised if at ~12,000 3dmarks its pretty much all CPU limited. I'm more interested in seeing ShaderMark scores.

 

[Dave Baumann]

The only test that has much impact from CPU performance is GT1.

 

[I.S.T.]

Um, that's only correct with 2001SE. 2003 is mostly vid card dependant. Edit: I was replying to nobie.

 

[bloodbob]

Wouldn't this most likely be a violation of the EULA??? after all the drivers haven't been certified for the NV40.

 

[volt]

As Ted mentioned I'm too, eager to know what drivers were used (if the inq. story is true at all)

 

[Bouncing Zabaglione Bros.]

1. Overclocked 9800 Pros are getting nearly 7000 3DMarks on 3DM2003

2. On NV3x, we saw Nvidia scores almost drop in half when the Nvidia "optimisations" were disabled, and then double again when Nvidia re-enabled them. Until we know if these are genuine scores or Nvidia cheats, such rumours are pretty meaningless.

This is just another Inq article that merely attempt to hit the side of a barn.

 

[Arun]

Keep in mind that due to a doubled register pool, the maximal advantages of the partial precision hacks in 3DMark03 are divided by (at least) 2.
Theorically, this means that if a NV3x had a score of 4000 with 52.16 and 5000 with newer drivers, a NV4x would in pure and simple THEORY need a score of 4500 with non-cheating drivers to get 5000 with cheatin drivers.

Keep in mind this is pure and simple THEORY. It is in no way truly representative of real differences between cheating & non-cheating drivers for the NV4x, because:
1) Legit optimizations are possible (although not many of them IMO, unless NV40's current compiler is godawful, or NVIDIA decides to use several hundreds manhours on optimizing 3DMark03 for the NV4x).
2) Other things than partial precision can be used to improve performance, although NVIDIA has already promised not to use them anymore, both to the public and privately to FM AFAIK.
3) The NV3+ chips (NV35, NV36, NV38) are not capable of FP32 MADs at full speed.
4) It's a different architecture (plus, saying the register file is "doubled" is not sufficient because several other, highly related things are also changed)


Uttar

 

[991060]

Uttar, do you mean the register file allocated to EACH pipeline?

 

[stevem]

It will be interesting to see which of the game tests gain most. It would be interesting if GT2&3 doubled in score... OTOH, if GT4 results double...

 

[nobie]

I know 3dmark03 is video card limited on todays cards. But on tomorrows cards that might not be the case...

 

[Unit01]

I'm really doubtful about those scores until someone provides some evidence to their "honesty". Couldn't think of a better word :/

 

[Tim]

Keep in mind that due to a doubled register pool, the maximal advantages of the partial precision hacks in 3DMark03 are divided by (at least) 2.

But nVidia also has heavy vertex shader "optimizations" for 3dmark in their current drivers. It is properly not a valid assumption that the use of partial precision is the only "optimization" for pixel shaders either.