The New and Improved "G80 Rumours Thread" *DailyTech specs at #802*

I think you have misunderstood me. Determining theoretical maximums is quite difficult, and I'm not actually that interested in doing so. The comparison was between the efficiency of the old vs. the new architectures, or, to what extent compromises were made to increase flexibility.

And just to confuse things entirely, let me switch back to something unrelated to GX2 :>

Code:

                   3DM05    3DM06    Bandwidth   ALU   core-clock
(8800GTX-GTS)/GTS:   11%      18%          35%   50%          15%

ALU number adjusts for #procs and relative speeds. If I adjust the 3d06 by subtracting 3k off for CPU, I get 25%. I believe I shall leave the interpretation up to the reader. ;/

 

your ALU number is wrong, it's not 50%: (8-6)/6 = 33% (exactly the bw increase )

 

"ALU number adjusts for #procs and relative speeds"

Assuming rumors are correct, the GTS is 1200Mhz and the GTX is 1350Mhz.

1350 * 128 - 1200 * 96
-------------------------------- = .5
1200 * 96

 

Am I the only one that sees a serious error in that equation?

 

Ops, my bad, I missed that line

 

I'm not understanding the implementation of their logic either.

 

it just computes how much more ALU 'power' has a GTX wrt a GTS, it assumes you can scale ALUs ops linearly..but we really don't know if that's true or not (there might be some interdependancy with other parts of the GPU we don't know about..)

 

Does G7X Pixel pipelines scale linear?

 

My guess looks good heh:

[(1350*96)-(1200*80)]/(1200*80) = .35

 

I think NVIDIA has defiantly benefited from the presence of well-established differentiated brands, not only by creating additional exposure via reviews and so on but also by creating zounds of "micro price/performance points" not addressed by the official line-up. Both are well-illustrated in the last batch of X1950 reviews - how many times have you seen "it would have been faster then Nvidia's offering if not for factory OC cards" used? Furthermore, I don't see how it created marketing confusion, since we know that even BFXGBGROMG 7900 GT IR F1 SLAYER GOLDEN SHOWER TURBO ANNIHILATOR MEITNERIUM EDDITION will perform at least as well as 7900 GT.

 

Latest news from german magazin pc-welt:
http://www.pcwelt.de/news/hardware/video/61081/index7.html

 

They also say its is 30% faster than X1950, on the first page.

http://www.pcwelt.de/news/hardware/video/61081/index.html

 

Can someone semi-translate what the article says?

 

Wish I could, I just used Google to translate it.

 

Says the same stuff that the table says which is the same as Vr-zone's original leaked specs, also they added in there 30% increased performance over the x1950xt, but that was with earlier drivers, new drivers should increase performance. But the performance numbers they are saying are more speculation.

 

roughly translated

then it goes on to say performance could further scale with DX10 etc etc...

nothing groundbreaking.

 

the article also says something about 700mio transistors - on 90nm (sounds very unlikely to me)

 

Yeah, me too.
Although i still remember hearing that 222M on a 130nm process back in 2004 was impossible, and then came NV40.

 

Here's some additional info:

NVIDIA unified architecture with GigaThreadTechnology
Full DirectX 10 ShaderModel 4.0 support
NVIDIA SLI Ready
16x full screen anti-aliasing (Lumenex Engine) [will work with HDR according to some sources]
True 128-bit floating point high dynamic-range (HDR) (Lumenex Engine)
NVIDIA Quantum Effects physics processing technology (Advanced shader processors architected for physics computation)
Two dual-link DVI (2560 x 1600 resolution displays)
NVIDIA Pure Video support
PCI-E Express support
Open GL 2.0 support
NVIDIA ForceWareUnified Driver Architecture
Built for Microsoft Windows Vista

Funny thing is that the GTX has 2 SLI connectors like ATI's X1950Pro and X1650XT. The GTS only has 1. This was also mentioned by one of the German sites linked before.

 

Umh..unified..I wonder what nvidia means with that.

SSAA FTW?

Not that games should care about this anyway, but I can see non realtime apps like offline
renderers taking advantage of it once they start to composite MANY HDR fragments together..