D9P/G94: 9600 GT on 19th of February

Errr, 9800 GTX is confirmed for February now? That's not what I last saw rumoured at least... Although I'd expect a near synchronous tape-out schedule here, so we'll see.

 

Considering that it seems that 9800-series is just G92's, there's no need for name tapeouts, unless new spins are made to try to reach higher clocks or something?

 

I have yet to see any rumour whatsoever mention that as more than speculation. There definitely has been some noise of that on the 9800 GX2 and it seems likely indeed, but the GTX is much more of a mystery at this point as far as I can tell.

 

The fact that GX2 uses it, and has so far been top of the line always on release, speaks towards GTX having G92 aswell.

 

Now:

-G92
-G94
-G96
-G98

Before:

-G80
-G84
-G86


Am i missing something, or is there an extra core in the refreshed Nvidia's lineup for H1'2008 compared to all of 2007 ?

 

It doesn't have to be...

Just check one of my previous posts. This is what one of my sources is currently saying:

D9E-40 - 9800GX2 (dual G92)
D9E-20 - 9800GTX (faster G92)
D9P-40 - 9800GT (slower G92)
D9P-20/10 - 9600GT (G94)
D9M-40/10 - 9500/9400 (G96)

See the overlap of certain cards and codenames? Could be a mistake... or it could be a big misinformation campaign... or it could be a big naming-mess....

G82 was canned.

G80 -> G92
G82 -> G94
G84 -> G96
G86 -> G98

GF9 series contains G92, G94, G96
GF8 series contains G80, G84, G86, G92, G98

So NV rebranded G92 to the GF9 Series, but will clock it much higher than a GF8800Ultra for the 9800GTX to justify the namechange.

 

What was the G82?

 

half the name change is justified by Hybrid Power, yea it doesn't look like much but I find that feature brilliant and it's what differenciates gf8 and gf9.

 

Hmm so it seems that GF9800GTX will be slower than GF8800GTX/GF8800Ultra in higher resolutions because of 256-bit memory bus and "only" (G80 has 24) 16 ROPs.

 

Nvidia 55nm Q2

Is the D9P-40 55nm?

Nvidia has 55nm chips tapped out!
http://www.fudzilla.com/index.php?option=com_content&task=view&id=5054&Itemid=1

In Q2 2008 D9P is 55nm?
http://www.digitimes.com/mobos/a20071129PD216.html

It say's D9P in June is that the D9P-40 55nm one?

 

Do you mean G92@55nm?? OK but the question is WHEN??

 

Sorry I havnt read all 9 pages to see if its allreay been posted but
its been benchmarked, Performance-wise, the 9600 GT delivers about 83% of the performance of the 8800 GT in 3DMark06, scoring 10357 3DMarks to the 8800 GT's 12508 3DMarks.
http://en.expreview.com/?p=184#more-184

 

First GF9600GT review!!

http://www.pconline.com.cn/diy/graphics/reviews/0801/1205247_5.html

 

Yep. 240mm² <-> 210mm²; seems the RV670 still has the smaller core, thanks to the 55nm production.

Interesting, especially when you think about it that Nvidia intends to use the same core for an even lower priced product (at least that's how I understood the discussiion here). So maybe we will see an even lower spec version of the HD38xx too simple because the cost structure of the RV670 should be even better here. Maybe because of that ATI/AMD didn't introduce a higher spec'd HD36xx.


Manfred

 

RV670 according to techreport is 192 mm^2

http://techreport.com/articles.x/13603

G84 is 169 mm^2

and taking dimensions on the thermal pad residuals is a little imprecise (in the photos you can see clearly that the instrument is not really touching the die)

I think G94 is a little smaller than 240 mm^2 (even if RV 670 is still smaller)

 

Remember that UMC has lower cost/wafer, but lower yields traditionally. At the same time, 55nm yields should also lag behind 65nm yields slightly, but I'm not sure the difference would be as substantial as between TSMC and UMC in general. And then again, yields don't mean much when you have potential redundancy nearly everywhere and that you can have plenty of bins to counteract parametric yield loss...

Overall, it should be fairly obvious that in terms of 'raw chip' cost (excluding package and testing which are harder to compare), RV670 is slightly cheaper than G94 (but probably by a bit less than the 240/210 ratio would imply). Where things get really hard to compare is when you factor in NV and AMD's different redundancy mechanisms, which both have advantages and disadvantages.

Also, these numbers are obviously better than those at expreview. CoJ and Crysis are especially good, and both are AFAIK quite ALU-heavy especially at high resolutions. This makes me suspect that we *are* looking at a higher ALU-TEX ratio than on G92 here; although whether that's via 96 SPs or a better config (consistent 3 flops/SP?) is a much harder question to answer.

P.S.: PCOnline consistently tends to overestimate chips' die sizes I think. My expectation is that the 'real 'die sizes are probably "~(192/210)*die size" for everything.