The G92 Architecture Rumours & Speculation Thread

[Arun]

It would appear NVidia's aiming to make two categories of SKUs from it, just like G80. Which begs the question, what size bus does the full chip have? Is it likely to stay at 256-bits, or will it have more ROPs/MCs and be 320/384-bits?

Well, the rumours including the leaked package/pin pictures seem to be indicating all SKUs will be 256-bit (or 192-bit in a few months if 1.1/1.2GHz GDDR4 becomes cheaper? Who knows!)

It seems strange to me that NVidia wouldn't release the fully-enabled chip before Christmas if the part-enabled chip is launching in about 1 month.

Same here, which is why I'm still presuming this is all just NVIDIA FUD and we'll see it before Christmas! On the other hand, I'd be surprised if we saw the 'GX2' this year.

Doesn't add-up...

Sure doesn't, which is why I'm forced to selectively ignore some illogical tidbits and add in my own speculation for all this to make any sense whatsoever in my eyes - however, I very well could be even more wrong because of this, who knows. Although I *am* rather confident here tbh!

 

[nicolasb]

It would appear NVidia's aiming to make two categories of SKUs from it, just like G80. Which begs the question, what size bus does the full chip have?

I thought the two SKUs were 256MB of memory and 512MB - difference in amount of RAM rather than bus-width.

nicolasb: 55nm is a 90% linear shrink, so the die size should be 19% lower than on 65nm. That would be 234mm2, thus, not 207mm2...

Are you saying 55nm is 90% of 65nm?

 

[Arun]

I thought the two SKUs were 256MB of memory and 512MB - difference in amount of RAM rather than bus-width.

Jawed is thinking of the 8800 GT as one 'SKU' here, and imagining that there is a higher-end model which hasn't leaked out yet for the sake of arguement.

Also, regarding 55nm being 90% of 65nm, yes, that's TSMC's definition - blame them, not me Process node naming is more about marketing than engineering nowadays anyway... http://www.xbitlabs.com/news/other/display/20070328080802.html

 

[nicolasb]

Jawed is thinking of the 8800 GT as one 'SKU' here, and imagining that there is a higher-end model which hasn't leaked out yet for the sake of arguement.

Wouldn't that be the alleged dual-PCB version due out in January?

 

[Arun]

Wouldn't that be the alleged dual-PCB version due out in January?

No, that would be a third SKU. We were arguing about single-chip solutions here.

I have some difficulty figuring out why 8 clusters/512MiB/825MHz/1.4GHz+ @ $399 is such a difficult thing to swallow. I know it's not what has been rumoured so far, so I'm not saying it's a certainty - but it does make a lot more sense than the alternatives, AFAICT.

I'm not saying it'll be a simultaneous launch though. I could imagine NVIDIA launching the 8800 GT in late October and then try to clear G80 stock for a month or so, then launch the higher-end SKU along with RV670 to steal its thunder. This is all speculation, of course.

 

[nicolasb]

Also, regarding 55nm being 90% of 65nm, yes, that's TSMC's definition - blame them, not me Process node naming is more about marketing than engineering nowadays anyway... http://www.xbitlabs.com/news/other/display/20070328080802.html

Note to self: stop expecting that anything remotely connected with the semiconductor industry will ever make logical sense.

 

[Sound_Card]

But we have heard of the second SKU based off the G92 core. Right? I thought the 8800GT was the second SKU based on G92.

 

[dnavas]

I have some difficulty figuring out why 8 clusters/512MiB/825MHz/1.4GHz+ @ $399 is such a difficult thing to swallow. I know it's not what has been rumoured so far, so I'm not saying it's a certainty - but it does make a lot more sense than the alternatives, AFAICT.

Does it dovetail with rumors of attempts to clock parts higher? Maybe the higher clocked part was always going to be a higher clocked part?

Do you think NV was targeting a family release originally? Having missed that, they moved the ready parts forward?

-Dave

 

[Arun]

Do you think NV was targeting a family release originally? Having missed that, they moved the ready parts forward?

That'd certainly make a fair bit of sense in my mind, yeah. I don't know if it's the case, but I wouldn't exclude the possibility.

 

[AnarchX]

I said and I'll say it again: I think the 8800 GT has 6 clusters activated and runs at 600/1800

Why should NV degrade it by a name below 8800 GTS, when it would be faster even then New-GTS in some/most? cases?

4 Clusters with a higher shader-clock-ratio like 1:3, would imo be enough to compete against 2950Pro(750/900MHz) and that its what NV probably wants.

6C@600MHz+1GHz+ memory, maybe come, when the G80 sold-out via New-GTS is done, as 8850 GTS or something else that indicates the higher performance.
That would make more sense or not?

 

[Arun]

The 8800 GT would have 10% lower bandwidth than the 8800 GTS, and less ROP power (unless the ROPs have been tweaked). Even with 20% higher core clocks, I'm not sure why it should be much faster than the current GTS, and I fail to see how it could be faster than the revamped GTS unless its architecture is significantly more efficient.

And the 8800 GT was originally called the 8700, as I'm sure you realize - NVIDIA's naming is not exactly very logical this round, unless they decide to change it at the last second again.

 

[AnarchX]

The 8800 GT would have 10% lower bandwidth than the 8800 GTS, and less ROP power (unless the ROPs have been tweaked). Even with 20% higher core clocks, I'm not sure why it should be much faster than the current GTS, and I fail to see how it could be faster than the revamped GTS unless its architecture is significantly more efficient.

texel fill-rate:
14.4GTex tri > 14 @ New-GTS > 12 @ Old-GTS
With higher AF-level bandwidth-difference should not be so important.

arithmetical power (@ Mobile1 was said ROP/TMUs low-speed but shaders high-speed, so I think it is a higher ratio than on 8800GTS/GTX, lets calculate with 2.47 like on Ultra -> 1.5GHz)

430GFLOPs > 403 @ New-GTS > 345 @ Old-GTS
Also somethings, thats not so much influenced by bandwidth.

The lower rop-power should also not so important, since G8x has extreme powerfull ROPs, which should not be limiting in most cases.

In my opinion your 8800 GT would be faster in most benchmarks than Old-GTS and New-GTS, and that would be confusing for customers and not very logical. (And think about factory-OC-versions of 8800 GT, which probadly will be sold <$300, which would make this circumstance more clearly)


Also we have to think about, why NV upgraded GTS?
In my opinion, they wanted to reach that "GTS" has a clearly right to exist, because 8800 GT came with its 9.6GTex tri/Pix and shader-power >300GFLOPs to near to the old version and AIBs would remain seated on this cards, because you can buy a 8800 GT for $249 while GTS costs $399 (maybe $349 in November), which is not really faster.

And that 8800 GT was called 8700 GTS in recent past, is also a sign imo, that it could not be faster than a 8800 GTS.


But lets see, maybe NV really gets illogical...

edit:
I just heard, that 8800 GT 512MB was reduced to $229.

 

[DegustatoR]

Also we have to think about, why NV upgraded GTS?

To compete with RV670?

 

[DegustatoR]

I just heard, that 8800 GT 512MB was reduced to $229.

And that kinda removes any sense from the 6-8 TCPs rumours.
Guys it looks like G92 is just 64 SP middle level card.
They'll update G80GTS to compete with RV670XTX.
Hi-end G9x should be ready in 1Q08 (or not if they want to counter R700 with G100 already).

 

[v_rr]

4 Clusters with a higher shader-clock-ratio like 1:3, would imo be enough to compete against 2950Pro(750/900MHz) and that its what NV probably wants.

You are very good guessing the future. Why don´t you open a store with astrologic previsions

Just to say, that guessing is not good.

edit:
I just heard, that 8800 GT 512MB was reduced to $229.

One more reason that G92 is slower than expected. The price tells you the performance, and if they reduced it´s price is because things ar as good as expected

 

[Sound_Card]

Nvidia may have just fumbled. If the 8800GT is a 64sp part at $229, it's going to have a really hard time with "Revival" which is exactly like R600 on a 256bit bus.

The "New" GTS is one akward SKU for sure. Still 90nm only means that it's going to be around $350 to $400 fighitng the $300 55nm "Gladiator" which is also faster than R600. Talk about a major margin avantage here too(55nm > 90nm --- 256bit > 320bit --- 512mb > 640mb).

 

[Vincent]

You are very good guessing the future. Why don´t you open a store with astrologic previsions

Just to say, that guessing is not good.


One more reason that G92 is slower than expected. The price tells you the performance, and if they reduced it´s price is because things ar as good as expected

Sorry, Have you ever tested the sample ???

 

[NonNative]

If I need to choose new 8800GTS and RV670XTX.

RV670XTX would wins this becuase of heat.

 

[PeterAce]

Well priced mainstream and performance-mainstream is all well and good..... but where is the 10 Cluster and 1.4 Ghz GDDR4 on a 384-bit Mem bus monster card to replace my 'almost a year old' 8800 GTX?

 

[AnarchX]

If I need to choose new 8800GTS and RV670XTX.

RV670XTX would wins this becuase of heat.

132W (2950XT) vs 147W(8800GTS) board-power(TDP).

Nvidia may have just fumbled. If the 8800GT is a 64sp part at $229, it's going to have a really hard time with "Revival" which is exactly like R600 on a 256bit bus.

Arithmetical power should be the key to beat RV670, tex/pix/z fil-rate have 4C/RPs @600MHz more than enough...