The G92 Architecture Rumours & Speculation Thread

[AnarchX]

I DONT UNDERSTAND.

Then take some time and read this:
http://www.beyond3d.com/content/reviews/1/

 

[Arty]

I'm starting to believe that the HD 3870 (god, what an awful name... :S) isn't the real competitor to the 8800 GT 512MB. Even the dual slot cooler doesn't add up.
The report that the manufacturers are already at work on non-reference PCB's for the 8800 GT is a further hint, seeing as the "true" high-end products are usually "shielded" by Nvidia from such practices.


So, i think a 128 scalar processor-sporting G92 with a dual slot cooler (higher clocks and, perhaps, a 320/384bit bus) may be unveiled down the road.

Well I've said previously that G92 = 8800GT does not add up, AT ALL. I also believe that G92 is a a full G80 with some tweaks. I'm not sure why Nvidia went ahead with such a naming convention (GT > GTS) .. its really stupid. It is one ball game to confuse your competitor but it does not mean you confuse your consumers! The 38xx nomenclature as awful as it may sound, looks like a better choice than anything both sides made in the last 5+ years.

My view was that AMD survived even without providing anything to match against the GTX AND the Ultra. Now that they have a cost advantage (& feature?) instead of a disadvantage (does that count as double jump? ), they are better positioned than the previous quarter.

 

[2900guy]

just wanna know if you guys think it will have 28 tmu. its a simple yes or no.

 

[Jawed]

So, i think a 128 scalar processor-sporting G92 with a dual slot cooler (higher clocks and, perhaps, a 320/384bit bus) may be unveiled down the road.

I agree (though I'm cagey about >256-bit bus).

The only fly in the ointment is my suspicion that NVidia has been struggling with 65nm ("A2" on die-picture at eXpreview doesn't really tell much though, does it?) and that G92 was supposed to be a "summer" part (i.e. it's late, as is G80's true successor, though I'm waiting till the fat lady sings on that latter detail before getting excited)...

The core/shader clocks on this 65nm part are bizarrely low in comparison with the 80nm G84.

But that doesn't rule out the possibility of a 128-SP 700 core 1800 SP part - just can't think what the SKU would be called though

Jawed

 

[INKster]

I'd just like to know exactly what is it with the 8800 GT's ALU's that gives them so much more performance than it was supposed to.
Don't tell me that those 16 extra ALU's have so much more performance that GTS' 96, since the GTX has an additional 16 on top of those and the speed difference is not nearly as wide.

It sports a memory bus cut down from 320bit to 256bit (not even the 1800MHz GDDR3 can make up for the fact that it has a few GB/s less overall than a GTS) and yet it kills the 8800 GTS 640MB precisely where it shouldn't according to the review.

The only fly in the ointment is my suspicion that NVidia has been struggling with 65nm ("A2" on die-picture at eXpreview doesn't really tell much though, does it?) and that G92 was supposed to be a "summer" part (i.e. it's late, as is G80's true successor, though I'm waiting till the fat lady sings on that latter detail before getting excited)...

The core/shader clocks on this 65nm part are bizarrely low in comparison with the 80nm G84.

How do you explain the 65nm 8400 GS sporting exactly the same clocks as the old 80nm version ?
And the fact that it came out at the same time as the 8800 GT/G92, despite the fact that it's a low-end, simple die-shrink (well, with the exception of the new VP3 processor anyway... ) ?
Or the fact that they've ordered a bunch of G92's (not G98/8400 GS 65nm) to UMC in addition to TSMC ?

The delay story doesn't make sense in light of these events to me.

 

[Arun]

just wanna know if you guys think it will have 28 tmu. its a simple yes or no.

Then you are on the wrong site.

 

[Jawed]

How do you explain the 65nm 8400 GS sporting exactly the same clocks as the old 80nm version ?

I wasn't aware of such a part. And, besides, you'd hope a much smaller die would clock faster...

And the fact that it came out at the same time as the 8800 GT/G92, despite the fact that it's a low-end, simple die-shrink (well, with the exception of the new VP3 processor anyway... ) ?

G92 hasn't launched yet...

Or the fact that they've ordered a bunch of G92's (not G98/8400 GS 65nm) to UMC in addition to TSMC ?

... and now we have rumours of limited supply for G92, so if NVidia went to UMC, perhaps it was as backup to problematic 65nm at TSMC.

The delay story doesn't make sense in light of these events to me.

Whereas it makes a lot of sense to me.

But then it would.

Jawed

 

[Jawed]

I'd just like to know exactly what is it with the 8800 GT's ALU's that gives them so much more performance than it was supposed to.

We've not had the benefit of B3D's penetrating analysis of why G84's ALUs are better than G80's - we're told to take it on trust.

But if they are, and if G92 uses the same ALU configuration as G84 (or as I mused earlier, the double-precision ALU configuration has a benefit of increased ALU capability in single-precision) then, Bob's your uncle.

Hey, Bob, you'll be around next week to tell us about the wonders of double-precision, won't you?

Jawed

 

[Arun]

I just realized that my "final guesses I won't update and won't think about much anymore until we're under NDA" got killed with the server crash, so here's a rough repost of those in case anyone interested missed them:

1x500-600MHz 6C; 12 ROPs/192-bit; 700-850MHz GDDR3; Q108; 384MiB @ $149+
1x600-600MHz 7C; 16 ROPs/256-bit; 900MHz GDDR3; Oct 29th; 256MiB @ ~$199; 512MiB @ ~$249
1x750-850MHz 8C; 16 ROPs/256-bit; 1400-1600MHz GDDR4; Q408; 512MiB @ ~$349; 1024MiB @ $399+
2x600-750MHz 8C; 16 ROPs/256-bit; 1100-1200MHz GDDR4; Q108; 512MiB @ ~$499; 1024MiB @ $599+

 

[nAo]

Yeah..but what about ALUs' frequency?

 

[AnarchX]

Why should have a 289mm² Die only 256Bit and 8 Clusters(32TMUs/128SPs)?

Also GX2 does not match with 3-Way-SLI and effective 512MB are to little for a 2008 high-end-SKU, like we see in the review at Expreview in the CoH and PT-Boats values.

 

[dizietsma]

It sports a memory bus cut down from 320bit to 256bit (not even the 1800MHz GDDR3 can make up for the fact that it has a few GB/s less overall than a GTS) and yet it kills the 8800 GTS 640MB precisely where it shouldn't according to the review.

Well the performance gap does decrease to about half as you put up the resolution but then I again I know what you are getting at.

I still find it very puzzling that the 8800GT is faster than the 8800GTS if, and only if, the 8800GTS continues to be sold.

Who in their right mind would by the more expensive GTS if it is that slower in general?

Perhaps they will dump the GTS and still bring out an 8700GT with 64/80sp.

 

[The Asgard]

Well the performance gap does decrease to about half as you put up the resolution but then I again I know what you are getting at.

I still find it very puzzling that the 8800GT is faster than the 8800GTS if, and only if, the 8800GTS continues to be sold.

Who in their right mind would by the more expensive GTS if it is that slower in general?

Perhaps they will dump the GTS and still bring out an 8700GT with 64/80sp.

There is a new GTS on its way.

 

[RobertR1]

Come on nvidia. Quit being lazy bastards and suprise us with a new high end Sku!!!

 

[ShaidarHaran]

There is a new GTS on its way.

Which is absolutely retarded. The GT nomenclature implies lower performance than the GTS. If there is a new GTS, it had better be named 8850, 8900 or some such.

 

[The Asgard]

Which is absolutely retarded. The GT nomenclature implies lower performance than the GTS. If there is a new GTS, it had better be named 8850, 8900 or some such.

Does seem crazy. 8900 GT would be far less confusing and appropriate for the performance if the leaks are anything to go by.

I suspect there is still lots of stock of the GTS.

 

[DegustatoR]

AFAIK G92 has 128 SPs and 256 bit bus, nothing more.
But NV is VERY good at keeping it all in secret these days.
So i won't be surprised if they have something bigger somewhere underground

 

[Domell]

What about ~800/2400MHz, with a consumption on 88GTX-level or a bit below like I heard?

Hmm is this info from reliable source and you`re talking about 8TCP GPU??

 

[_xxx_]

I'll make one of my crystal ball unfounded out of thin air predictions again: 192 SP, 512-bit bus and 1 GB mem for the next nV high-end part (next as "this round") trouncing AMD/ATI into oblivion. Big and bold.

 

[Geo]

If G92 is 512-bit to external memory for one gpu then I'll borrow a hat to eat it (I like my own hats too well for that).