NVIDIA GT200 Rumours & Speculation Thread

[tacopaco]

So, other than NV30 and NV50(G80), it is clear that NV tends to transition to a new architecture in 13-16 months on average, with the most recent GT200 only slightly outside this range @ ~18 months.

right they've followed a distinct pattern over the years and if the pattern was disrupted it was because the architecture did better than expected (g80) or worse 6 series->7

 

[ninelven]

increase as AMD pushes their own cost effective multi-GPU solutions.

Being "cost effective" doesn't really help so much as being profitable. I probably have a gloomier outlook than most on the apparent multi-GPU strategy though.

 

[trinibwoy]

Being "cost effective" doesn't really help so much as being profitable. I probably have a gloomier outlook than most on the apparent multi-GPU strategy though.

Profit is a simple function of cost and price. I'm not sure how you could just toss the former out the window.

 

[ninelven]

Well, when you can charge whatever you want for your product that helps . Selling a lower volume of an (relatively) inexpensive to produce product at lower profit margins OTOH....

Profit is a simple function of cost and price.

If you ignore R&D and actually paying employees...

 

[tacopaco]

actually its profitable especially with the 3870x2 two gpu on one pcb saves production costs and they can sell it for a high profit margin and it would still be "cost effective" especially when compared to similar solutions from nv

 

[tacopaco]

^^^^ at the time*

 

[ShaidarHaran]

Unfortunately we're in interpretation-soup - you could just interpret this as execution of fp32 texture filtering in the ALUs using texels fetched by the TMUs - hmm, G80 can do fp32 texture filtering can't it?

I haven't read it at all closely, just perused the diagrams.


I agree. Which is why I'm willing to accept it as fully implementable as software at some point. And also why not every unit described may be implemented in current hardware - instead units might be amalgamated.


An old theory relating to int8 rate was simply register pressure and I still think this hasn't been fully explored.

Separately, did you look at the floating point filtering patent document? This is clearly relevant to the fp10 and fp16 cases at minimum, I'd say. EDIT: in fact it's quite explicit about doing fixed point, "s2.14" filtering, i.e. 16-bit - so it would seem it can't do fp32 filtering and so fp32 filtering would have to be performed on the ALUs.

Jawed

If I understand correctly, this would mean that the function of the TF units has been reassigned to the SPs while the TAs remain (functionally) unchanged.

No, the key difference is R580 just upped the ALU's but it was texture bottlenecked to hell so it didn't do a vast amount of good..

Ah, but you see the increase in ALUs was very necessary, as R520's bottleneck was almost always with the ALUs. Providing more ALUs eliminated this bottleneck, which means the next bottleneck in line took its place (thus the chip became texture-bottlenecked).

G80 already has vast texturing power (and GT200 maybe a bit more), so just doubling the shaders could bear double perfomance in this case.

I agree. G8x/G9x products certainly have an abundance of texturing capability. Upping the SP count alone could provide near-linear performance increases (assuming the aforementioned setup bottleneck has been alleviated).

And double performance is of course roughly what major chip revamps aim for.

Incorrect. SKUs in new architectures aim for approximately double the performance of their counterpart first generation products. Refreshes usually only bring 15-30% more performance on average.

edit: then again, I'm not sure exactly what qualifies as a "major chip revamp" in your mind since that's not official nomenclature...

As to GTx2 being impossible .. impossible[!] ... i will disagree with you and then shut up about it. GTx2 is nothing i can prove and i will not attempt it. If you google "GT200x2" there is a lot of speculation about it and i guess i started that speculation by myself at ATF months ago. Maybe i will live it down; it is also not impossible i could also be right - especially next year, when there is a real option to make GT200 into a sandwich - then with the shrink and speed-bump; not now.

It's not all together impossible, but it is extremely cost-prohibitive both from a design and manufacturing standpoint. Dual 1GB 512-bit monolithic dies (even utilizing NV's dual-PCB design) is just too much for 65nm, even 55nm. The G70->G71 transition proved this. NV waited to release the dual-GPU 7950GX2 until the shrink to 90nm (from 110nm) had been completed. Also, NV utilized mobile (read: optimized strictly for power consumption), clock-reduced GPUs instead of standard 79xxs.

The 9800 GX2 didn't go quite that far (not using mobile parts), however it still uses lower-clocked and power-optimized GPUs than the desktop single-GPU counterpart (8800 GTS 512/9800 GTX).

i bet they will make one just because of their history 7950,9800gx2 i think they will introduce it late it its life just like the others were introduced end of g70,g80-g92 i remember reading somewhere that they see much of a future for sandwiched card and that a powerful single card is better than a sandwiched card so i see sandwiched cards as a means for them to squeeze all the money they can out of that architecture (after all it does cost millions to develop it) so i never buy the sandwiched cards because i have the patience to wait for the new architecture which usually destroys the power hungry sandwich cards anyway

If there is to be a dual-GPU GT200, it will only be a shrunken derivative thereof, not GT200 itself. This is for reasons already mentioned.

My point is that the incremental cost to go monolithic for them has actually gone *down*. And Lukfi, I'm not aware of NV ever having said the GF7 Series was a mistake; management loved that series, it had awesome margins, awesome ASPs, awesome market share.

And sorry, but for this kind of thing, Jen-Hsun is honest the vast majority of the time; sometimes he's simply wrong or misinformed, but he has better things to do than to lie outright to investors. I've listened to more than enough NV CCs to know he's not the kind of guy who'll try to downright confuse everyone even if it's in his interests. He got a PR and marketing team to do that, though.

Arun, I think you're misinterpreting Jen-Hsun's statements slightly. I do not read his answer to Hans Mosesmann's question in the same way. I believe he is simply stating a preference for monolithic designs, not ruling out SLI-on-a-card solutions in the future.

right they've followed a distinct pattern over the years and if the pattern was disrupted it was because the architecture did better than expected (g80) or worse 6 series->7

GF6 -> GF7 wasn't considered an architectural change though, since it was basically just a die shrink with minor architectural tweaks (G70-G71 was an even lesser refresh with simply a die shrink and increased clocks). It sucks to have that many refreshes in a row though, at least from an enthusiast's perspective.

 

[tacopaco]

i didn't realize it was such a minuscule change i had just assumed gf7 was a refresh as for g70-g71 i knew it wasn't a major change
as for enthusiasts it sucks but they also have to knowledge to pick and choose which refresh is in their best interests that gives them the upper hand over mainstream consumers

 

[tacopaco]

If there is to be a dual-GPU GT200, it will only be a shrunken derivative thereof, not GT200 itself. This is for reasons already mentioned.

i understand that and already was aware that it would be a derivative 40nm because current manufacturing processes are to big for two gt200 cores on one chip

 

[trinibwoy]

Well, when you can charge whatever you want for your product that helps . Selling a lower volume of an (relatively) inexpensive to produce product at lower profit margins OTOH....

If you ignore R&D and actually paying employees...

Heh, I'm not sure what R&D and SG&A costs have to do with a product strategy.

Lower manufacturing costs are not in any way related to or indicative of low volumes or lower profit margins. As a matter of fact they are a key ingredient in higher profit margins - the other major factor being higher ASPs of course. They are equally important so I'm not sure why you think one is more important than the other....

 

[tacopaco]

Lower manufacturing costs are not in any way related to or indicative of low volumes or lower profit margins. As a matter of fact they are a key ingredient in higher profit margins

thats what i was saying earlier

 

[ShaidarHaran]

i understand that and already was aware that it would be a derivative 40nm because current manufacturing processes are to big for two gt200 cores on one chip

Perhaps it was someone else then... My mistake.

 

[ninelven]

Lower manufacturing costs are not in any way related to or indicative of low volumes or lower profit margins.

I never said this... that would be a gross generalization. I was merely commenting on a (individual) potential reality.

As a matter of fact they are a key ingredient in higher profit margins

It can be... that is not always the case though.

Heh, I'm not sure what R&D and SG&A costs have to do with a product strategy.

Well, generally companies like to stay in business so that they can actually have future products much less strategies...

 

[trinibwoy]

I never said this... that would be a gross generalization. I was merely commenting on a (individual) potential reality.

Understood but the potential reality you were referring to is one in where AMD has both lower costs and lower ASPs which isn't necessarily going to be the case. It's quite possible that their multi-GPU technology matures to the point where they can again push products into the $500-$650 segment while keeping costs lower than the competition.

Well, generally companies like to stay in business so that they can actually have future products much less strategies...

Yeah but how does the monolithic vs multi-GPU decision affect R&D or other costs? Neither approach is inherently more expensive as there are a lot of different factors at play.

 

[Megadrive1988]

If I understand correctly, this would mean that the function of the TF units has been reassigned to the SPs while the TAs remain (functionally) unchanged.



Ah, but you see the increase in ALUs was very necessary, as R520's bottleneck was almost always with the ALUs. Providing more ALUs eliminated this bottleneck, which means the next bottleneck in line took its place (thus the chip became texture-bottlenecked).




I agree. G8x/G9x products certainly have an abundance of texturing capability. Upping the SP count alone could provide near-linear performance increases (assuming the aforementioned setup bottleneck has been alleviated).



Incorrect. SKUs in new architectures aim for approximately double the performance of their counterpart first generation products. Refreshes usually only bring 15-30% more performance on average.

edit: then again, I'm not sure exactly what qualifies as a "major chip revamp" in your mind since that's not official nomenclature...



It's not all together impossible, but it is extremely cost-prohibitive both from a design and manufacturing standpoint. Dual 1GB 512-bit monolithic dies (even utilizing NV's dual-PCB design) is just too much for 65nm, even 55nm. The G70->G71 transition proved this. NV waited to release the dual-GPU 7950GX2 until the shrink to 90nm (from 110nm) had been completed. Also, NV utilized mobile (read: optimized strictly for power consumption), clock-reduced GPUs instead of standard 79xxs.

The 9800 GX2 didn't go quite that far (not using mobile parts), however it still uses lower-clocked and power-optimized GPUs than the desktop single-GPU counterpart (8800 GTS 512/9800 GTX).



If there is to be a dual-GPU GT200, it will only be a shrunken derivative thereof, not GT200 itself. This is for reasons already mentioned.



Arun, I think you're misinterpreting Jen-Hsun's statements slightly. I do not read his answer to Hans Mosesmann's question in the same way. I believe he is simply stating a preference for monolithic designs, not ruling out SLI-on-a-card solutions in the future.



GF6 -> GF7 wasn't considered an architectural change though, since it was basically just a die shrink with minor architectural tweaks (G70-G71 was an even lesser refresh with simply a die shrink and increased clocks). It sucks to have that many refreshes in a row though, at least from an enthusiast's perspective.

GF6 -> GF7 indeed was not a total architectural change, but was a significant refresh. Not just a die shrink+tweak
as G70 -> G71 was.

 

[Megadrive1988]

I think it was Jen-Hsun himself who said GF7 was a mistake, it was a kinda semi-official statement.

No he said GeForce FX (NV30 / GF5 ) was a mistake, not GF7.

 

[Megadrive1988]

Is there a G80 GX2?

Not exactly. the 9800 GX2 is two G92s. The G92 is a shrunk, streamlined version of the G80.

 

[no-X]

Reread Jen-Hsun's reply, he clearly hints at the 9800 GX2 there. And Lukfi, I am absolutely positive that NVIDIA always loved and still loves the GF7 series. In fact they initially expected a substantially slower transition than what actually happened (i.e. they expected OEMs to stick to GF7 for a variety of designs in the 2007 Back-to-School cycle, but nearly every design switched to G84/G86 or RV610/RV630)

I'm 100% sure that nVidia (and probably Jen-Hsun) marked GF7 as a mistake. I'll try to find it.

//ed: Jen Hsun said it to press at Santa Clara before launch of G80 (October 2006 - fixed typo, thx to AlphaWolf).

 

[apoppin]

You can only edit after a certain number of posts and days, see: http://forum.beyond3d.com/showthread.php?t=23045
As for GT2xx GX2; it's not that they can't, they just won't have to because they'll go monolithic every single generation from now on.

i believe that you really mean, nVidia expects to stay monolithic and that they would love to have a single powerful core that dominates all.
-But then, they really don't know what performance r700x2 [x3/x4] might bring, and they may be forced to make a GT200-X2 [variant].

Maybe; not impossible. And the only thing i can do now is to bet you. But betting is illegal in the USA. A gentleman's wager, perhaps?
[you don't need to point out i am no gentleman; but i am certain there will be a GT200x2 on the smaller process next year!
- Edit: How certain am i? i will apologize to all of you and leave B3D in shame forever; -- even if i somehow make it as a member here to 01/01/10 =P]

If I remember correctly, they were GF9800 specs that someone copied from some Asian forum (in native language) and didn't bother to translate the line that said that it's just the poster's personal theory. Even though someone in the thread pointed out (more than once) that those specs are not real, those posts went completely unnoticed by the hordes of drooling users. (It may sound racist, but while reading VR forums, I can't help thinking there must a be a horrible IQ difference between Asians and the rest of the world.)

The differences are cultural, imo. And some of the English speaking Tech sites do exactly the same, Lukfi.

 

[AlphaWolf]

//ed: Jen Hsun said it to press at Santa Clara before launch of G80 (October 1996).

That date has got to be wrong. 2006?