NVIDIA Fermi: Architecture discussion

[Razor1]

I find your G80 part deux very hard to believe.

yeah its hard to believe, but effective flops on fermi is 2.5 to 3 times that of the gt200. 633 gflops vs 1.5 -1.8 tflops.

if we look at the gt 200b, its ~2.5 times more

But wasn't talking about performance anyways, die size I haven't seen anything solid at all outside of it might be around these numbers.

 

[ShaidarHaran]

yeah its hard to believe, but effective flops on fermi is 2.5 to 3 times that of the gt200. 633 gflops vs 1.5 -1.8 tflops.

if we look at the gt 200b, its ~2.5 times more

But wasn't talking about performance anyways, die size I haven't seen anything solid at all outside of it might be around these numbers.

Wait - I thought they rectified the MADD/MUL co-issue with GT200 and FLOPs were at 900-1000 (depending on clocks).

 

[Ailuros]

Just need a little clarification on your post. The clockspeeds you stated you mentioned them regarding the target clockspeeds, correct? Not the A2 clocks.
Thanks.

Yes. But I have severe doubts that if A1 was already at 495/1100, that A2 ended up at only 500/1200 as some indicate. I think some just get a bit confused with Tesla frequencies. Afaik frequencies on final SKUs will scale as follows from higher to lower:

GeForce--> Quadro--> Tesla

So, you're saying that no one knows how big the chip is that they have to design their boards for?

Depends where someone starts to measure and there are always some differences in the end. GT200@65nm was supposed to be 575 and we concluded in the end that it's 583mm2 here at B3D I think. From what I've heard GF100 is 22.3*22.3, but I'm not laying my hands into fire for it either. But then again I separate myself from hearsay and knowledge.

 

[neliz]

But then again I separate myself from hearsay and knowledge.

Let's say it was hearsay first and became knowledge last week.

 

[mboeller]

Let's say it was hearsay first and became knowledge last week.

What size?

22,3 * 22,3 or 23,x * 23,x ??

 

[Ailuros]

Let's say it was hearsay first and became knowledge last week.

Knowledge in the strict sense means you've seen it with your own eyes.

 

[CarstenS]

This is probably more for the 'signs of strain' topic, but do you reckon that the margins in the consumer high end really matter that much to Nvidia? It's a small volume sector, and Nvidia makes no secret of the fact that most of their profit comes from the professional sector - where margins are still pretty strong.

Yes, I do. FWIW, Nvidia and it's partners had to sell their lower-high-end part as of late at a middle class price point, which has led to middle class products being sold even lower and so on.

I've never seen the space from 120 euros downward so cramped with quite differently performing products - that cannot be healthy in the long term for a regular company.

 

[Silus]

GT300 is the same size as GT200, i.e. ~24x24. Any lower (like speculated by some here since before may) is unfounded. Btw, You promised us pics of Fermi back then, still waiting for your sources to show any signs of credibility.

So it's 576 mm2 ? I have some doubts about that one...
Tech-Report speculates ~470 mm2- Jawed speculated about the same (a bit larger) and rumor sites (except of course the "notaccurate" one) all mention ~500 mm2, so around the size of GT200b, not GT200.

And what do you mean "I" promised pics of Fermi ?

 

[Razor1]

Wait - I thought they rectified the MADD/MUL co-issue with GT200 and FLOPs were at 900-1000 (depending on clocks).

no the dual issue never was rectified.

 

[ShaidarHaran]

So then how do they get away claiming such inflated SP GFLOP numbers?

 

[Razor1]

in gpgpu applications the dual issue was used more.

 

[DegustatoR]

My predictions are not good?

I for one am still waiting for that ressurected GT212 which you promised we'll see this year.

 

[neliz]

Tech-Report speculates ~470 mm2- Jawed speculated about the same (a bit larger) and rumor sites (except of course the "notaccurate" one) all mention ~500 mm2, so around the size of GT200b, not GT200.

I take my sources accuracy over their guesstimates any day.

And what do you mean "I" promised pics of Fermi ?

That wasn't directed at you, but at the people that have spent the last half year claiming the weirdest things, we'd have pictures in May, that it would be with us "soon" and for some magical reason shrinking from 55 to 40nm would magically double performance while also being a smaller die size, it does not compute.

I for one am still waiting for that ressurected GT212 which you promised we'll see this year.

That was layed to rest a few days later. But I'll give it to you as soon as you show me your GT300 pics from back in May.

in the meantime shouldn't we start to get worried about heat now the display connector setup is the same as on the RV8x0 cards?

 

[rpg.314]

in gpgpu applications the dual issue was used more.

Care to mention even one? The way I see it, zero>zero isn't right.

 

[Razor1]

I take my sources accuracy over their guesstimates any day.

unless your sources are people that made the chip, no one knows the size right now , that info was never released at least not yet.

That wasn't directed at you, but at the people that have spent the last half year claiming the weirdest things, we'd have pictures in May, that it would be with us "soon" and for some magical reason shrinking from 55 to 40nm would magically double performance while also being a smaller die size, it does not compute.

Kinda confused there, if its a different architecture why not? You aren't thinking about the increased cache sizes, the l2 cache, those take up transistors too, and are very packed much more dense. The only way I see how you came up with the die size is if you calculated from the gt200b die and transistor amounts, and crazily it comes out to the size you are pointing at. I don't think that is accurate.

 

[Razor1]

Care to mention even one? The way I see it, zero>zero isn't right.

the programming guides give you how to use the dual issue in the g80 and up.

 

[neliz]

unless your sources are people that made the chip, no one knows the size right now , that info was never released at least not yet.

How come you know nVidia didn't update it's partners and engineers the past few weeks?

 

[Razor1]

How come you know nVidia didn't update it's partners and engineers the past few weeks?

you have been saying that size for a while. And I just asked around this morning.

 

[Sontin]

How come you know nVidia didn't update it's partners and engineers the past few weeks?

They speak about the die size but not about the clocks, features and performance?

 

[CarstenS]

Care to mention even one? The way I see it, zero>zero isn't right.

Does GPUBench count?
http://graphics.stanford.edu/projects/gpubench/results/

Compare this to 8800 GTX there

--------------------------------------------------------------------------
Video Driver Information
--------------------------------------------------------------------------
GL_VENDOR: NVIDIA Corporation
GL_RENDERER: GeForce GTX 280/PCI/SSE2
GL_VERSION: 2.1.2
Driver Version: 7772

[...]

--------------------------------------------------------------------------
Instruction Issue
--------------------------------------------------------------------------
512 70.1201 ADD 4 64
512 70.1401 SUB 4 64
512 94.9697 MUL 4 64
512 68.8565 MAD 4 64
512 143.0448 EX2 4 64
512 72.1732 LG2 4 64
512 68.0586 POW 4 64
512 136.2859 FLR 4 64
512 68.6366 FRC 4 64
512 70.7210 RSQ 4 64
512 143.7839 RCP 4 64
512 143.0575 SIN 4 64
512 141.5776 COS 4 64
512 137.1915 SCS 4 64
512 228.6909 DP3 4 64
512 158.8046 DP4 4 64
512 89.4942 XPD 4 64
512 69.8814 CMP 4 64

--------------------------------------------------------------------------
Scalar vs Vector Instruction Issue
--------------------------------------------------------------------------
512 289.1751 ADD 1 40
512 70.5597 ADD 4 40
512 282.9698 SUB 1 40
512 69.9241 SUB 4 40
512 385.1988 MUL 1 40
512 95.8897 MUL 4 40
512 279.8798 MAD 1 40
512 68.0741 MAD 4 40

in the meantime shouldn't we start to get worried about heat now the display connector setup is the same as on the RV8x0 cards?

Pics? pah-leeeease. The last "SLI-Fermi" picture seems not to prove your point (albeit not denying it either)