55nm G92b die-shot @ PCPer

Doesn't PCPer actually have the card? They should be able to provide a somewhat accurate measurement.

 

And they did - 231 mm².

 

Well based on the coin shot, it does look like a 20-21% scaling to me, resulting in a 260-265mm² chip instead of a 330mm² chip. Although from another POV, the coin shot makes no sense either: the coin is slightly wider than it is tall, while based on the fact it's diagonal on the horizontal direction (if you see what I mean...) it should be slightly taller instead.

This is a bit ridiculous, having a ruler-based measurement and a die shot with a coin next to it, and we can't even decide what the die size is precisely. Bah!

 

The 2 coin shots are really the same scale, and both around 1.6% wide. By upscaling y i get 183x188 pixels for 65nm and 164x168 for 55nm, ie a 20% reduction in area. I guess he hasn't measured the 65nm for a comparable number...

 

GX2 is EOL. The longer it is out there, its just cannibalizing GTX280/260 sales.

 

G92b is already going to cannibalize GTX260 sales, IMO.

Why would you pay twice as much for the GTX 260 when it has such a small performance boost? It has 22% lower texture rate and only 2% higher math speed. The only plus is 75% more BW and the architectural tweaks.

In the end, a GTX 260 will be maybe 25% faster with AA enabled. It won't even buy you one resolution step.

 

Seems like someone did a bad job with these 2 screenshots.

I will try to do better...


First, this is my "old" die sizes comparison (now with R600) :




My "old" G92b size estimation is based on the TSMC documentation:

I used my photoshop skills to adjust these 2 new screenshots. ( Sorry for the oversized picture )




An now some maths based on this animate GIF :

G92 ...: 18mm x 18mm = 324mm²

G92 ...: 112 pixels x 111 pixels = 12432 pixels
G92b .: 102 pixels x 99 pixels = 10098 pixels

Ratio .: 12432 / 10098 = 1.2311 ---> 324mm² / 1.2311 = ~263mm²

G92b .: 16.39mm * 16.05mm = ~263mm²



My "old" estimation : ~270mm² (TSMC doc)
My new estimation : ~263mm² (Screenshots)
Difference : 2.5%

 

Heh. I actually came up with ~272mm^2 with these new pics by counting pixels.
But in any case, looks like it's pretty much the same size as rv770... So those now really should have the same manufacturing costs (save yield issues depending how redundancy is handled), maybe this should help answer who's got the more efficient, power-efficient etc. design .

 

Let's see...

Code:

                    [B]Core      Shaders    Memory     Size[/B]
[B]GeForce 9800 GTX+[/B] : 738 MHz   1836 MHz   2200 MHz   270mm²
[B]Radeon HD 4850[/B]    : 625 MHz    625 MHz   1986 MHz   256mm²
[B]Difference[/B]        :  -18.1%    -293.8%     -10.8%    -5.5%

Performance: Same without AA, and HD 4850 will be faster once AA and resolutions go up.

So the more efficient/mm² GPU is : ____________

 

Well, I'd say the G92b keeping up with something that has 2x the FLOPs is fairly impressive...

 

but there's a still G94b in the works (and even G96b). a 55nm 9600GT+ (w/ hybrid sli) at 100 euros could be an interesting proposition where cost and power are a concern. of course G92 always has been too cheap for G94 to shine (and now there's RV770 which is damn too cheap)
It will be used in laptops anyway. (and G96 is almost a laptop-only chip? I never found a 9500GT card for sale)

 

You've seen a 9500 in a laptop? I was under the impression they weren't even shipping them yet, because I certainly have never seen it.

 

PUMA + RV770 + SB750 + XGP (eXternal Graphics Platform) =

 

now I think of it, there's 9500M GS but it's actually G84. but I found an OEM PC with 9500GT, so thanksfully my impression of nvidia designing GPU and not selling them is off .

 

http://www.firingsquad.com/hardware/amd_radeon_hd_4850_geforce_9800_gtx+/

 

According to BitTech:

http://www.bit-tech.net/news/2008/06/19/nvidia-has-9800-gtx-55nm-g92/1

G92 - 324 mm^2 (18x18 mm)
G92b - 289 mm^2 (17x17 mm)

So only ~11% shirnk!

 

...and RV770 keeping up with something that has 18% more fillrate, 89% more texture rate, 136% more Z rate, and 10% more BW is also fairly impressive.

BTW, RV770 has only 42% more flops.

 

A GPU with that kind of power draw is definitely a better candidate for XGP than for being built in.
While there are merits to the XGP idea, I feel that it makes most sense as a means to add decent graphics to laptops with very weak IGPs. I fear that the manufacturers of said laptops can be hesitant to promote expansion capabilities that take away an argument for an up-sell.

 

Maybe they cut down some parts

 

Sure, if you count marketing numbers.