Nvidia BigK GK110 Kepler Speculation Thread

[silent_guy]

Well, unfortunately they will release GeForces with GK110 only in 2013.

Read more: http://vr-zone.com/articles/nvidia-s-monster-gpu-for-tesla-k20-2013-geforce-and-quadro-cards/15884.html#ixzz1v6ePIBpA

But, does it mean we will see GK114 as GTX 760 while a little bit later GK110 as GTX 780?

Which leads to the observation that Nvidia may find itself in a very comfortable situation for GeForce, doesn't it?

After discounting for DP support in 7970, its current architecture efficiency seems to be similar to the one of AMD, unlike the Fermi generation. I don't expect that either company can squeeze *that* much more performance per mm2 or performance per W out of 28nm. So whatever AMD comes up with for its next generation, Nvidia already has an answer waiting in the wings and can counter it and outperform it simply by enabling just enough SMs as necessary to take the crown. Reserve the good dies for high margin Tesla, recoup the remaining ones for GeForce. It's a strategy that can work if Tesla volumes are high enough.

I expect that AMD won't have a choice but to go very big die eventually.

 

[Ailuros]

And according to several other sites like http://www.anandtech.com/show/5699/nvidia-geforce-gtx-680-review/2

Well maybe Damien Triolet can shed some light into that one.

 

[AnarchX]

GK110 Whitepaper:
http://www.nvidia.com/content/PDF/kepler/NVIDIA-Kepler-GK110-Architecture-Whitepaper.pdf

SMX: 192 single‐precision CUDA cores, 64 double‐precision units, 32 special function units (SFU), and 32 load/store units (LD/ST).

When these units are physical, why they are not exposed to be used for SP processing?

 

[kalelovil]

Which leads to the observation that Nvidia may find itself in a very comfortable situation for GeForce, doesn't it?

After discounting for DP support in 7970, its current architecture efficiency seems to be similar to the one of AMD, unlike the Fermi generation. I don't expect that either company can squeeze *that* much more performance per mm2 or performance per W out of 28nm. So whatever AMD comes up with for its next generation, Nvidia already has an answer waiting in the wings and can counter it and outperform it simply by enabling just enough SMs as necessary to take the crown. Reserve the good dies for high margin Tesla, recoup the remaining ones for GeForce. It's a strategy that can work if Tesla volumes are high enough.

I expect that AMD won't have a choice but to go very big die eventually.

But while Nvidia will have spent their next-generation R&D resources on making a workable GK110, which will in all likelihood be less efficient in gaming workloads than GK104, AMD will be able to devote it to a proper new generation.

Being stuck on 28nm for a while will limit the generational leap, but GCN is a largely new and innovative architecture which generally means there will be significant low-hanging improvements to be made to both gaming and GPGPU workloads with little die cost.

Remember also that there is supposedly a lot of redundancy built into Tahiti to cope for an immature 28nm process and which won't be needed in the future.

 

[Ailuros]

GK110 Whitepaper:
http://www.nvidia.com/content/PDF/kepler/NVIDIA-Kepler-GK110-Architecture-Whitepaper.pdf




When these units are physical, why they are not exposed to be used for SP processing?

Even better: accepting that reasoning (which I still believe is only schematic but would like to stand corrected) would mean that GK110 doesn't have only 2880SPs but 3840 in total.

 

[UniversalTruth]

But while Nvidia will have spent their next-generation R&D resources on making a workable GK110, which will in all likelihood be less efficient in gaming workloads than GK104, AMD will be able to devote it to a proper new generation.

It will be a very interesting comparison when GK110 with ~3000 SPs (double ! that of GK104) comes with at least 50% more performance than GK104.
Meanwhile, don't expect more than 15-20% performance improvement over Tahiti for the next AMD one.
BTW, this one should have already been taped out.

Remember also that there is supposedly a lot of redundancy built into Tahiti to cope for an immature 28nm process and which won't be needed in the future.

Yeah, because at the same time there is nothing hidden in GK104.

Although, I agree that if AMD comes with a Barts-type die at 250-300 with all ridiculous GPGPU stuff removed, then they have some chances against GK114.

 

[kalelovil]

Yeah, because at the same time there is nothing hidden in GK104.

Based on its poor yields, there isn't as much redundancy as there needed to be.

 

[Alexko]

But while Nvidia will have spent their next-generation R&D resources on making a workable GK110, which will in all likelihood be less efficient in gaming workloads than GK104, AMD will be able to devote it to a proper new generation.

What do you mean by less efficient?

 

[Mulgogoi]

GK110 Whitepaper:
http://www.nvidia.com/content/PDF/kepler/NVIDIA-Kepler-GK110-Architecture-Whitepaper.pdf

Polymorph engine isn't shown in SMX when I see block diagram.

Does it mean that GK110 will be used only for computation purpose?

 

[silent_guy]

But while Nvidia will have spent their next-generation R&D resources on making a workable GK110, which will in all likelihood be less efficient in gaming workloads than GK104, AMD will be able to devote it to a proper new generation.

Being stuck on 28nm for a while will limit the generational leap, but GCN is a largely new and innovative architecture which generally means there will be significant low-hanging improvements to be made to both gaming and GPGPU workloads with little die cost.

Remember also that there is supposedly a lot of redundancy built into Tahiti to cope for an immature 28nm process and which won't be needed in the future.

Will GK110 be less efficient at games? Why? And why does it even matter, if you can compensate for that by selling Teslas? It didn't seem to be a problem during the Fermi generation, yet there the efficiency difference was substantial.
Does Tahiti have a lot of redundancy? Where?
Is AMD the only one with low hanging fruit? SMX is not exactly a carbon copy of the Fermi SM either.
What makes you think that Nvidia defect rates/yields are worse than those of AMD? Is there any reasonable argument to assume this would be so?

 

[Ailuros]

Depends what someone means with "less efficient than GK104 for gaming". The average user won't notice much of a difference since GK110 desktop should end up on estimate about 40-50% faster than GK104.

The nasty detail after that for those that have a little bit of understanding on those things is that the transistor and die area difference between GK110 (7.1b/550mm2?) and GK104 (3.54b/294mm2) is by quite a bit higher than if you compare GF110 (3.0b/529mm2) and GF114 (1.95b/365mm2).

On the other hand if you want to play the AMD-NVIDIA comparing game (which I'm not particularly fond of) 560Ti was from what I recall roughly 25% behind the 6970, while the difference between 6970 and 580 was roughly in the 15% league in favor of the latter. This time the actual performance part from NV (GK104) is a tiny bit ahead of Tahiti, which simply means that a lot of factors have changed this time.

Of course had/has AMD quite a bit of time until Q4 where GK110 is projected for, but on the other hand it doesn't seem to me all that likely that they'll manage to break even with the GK110 by then with a refresh part let alone surpass it. That's something different price points will take care of at the time anyway and I wouldn't hesitate to expect that AMD might have another price/performance winner against GK110 until then.

 

[DSC]

Full GTC 2012 keynote video posted on Youtube. Just linking part 1, watch the other parts from Nvidia's Youtube channel.

http://www.youtube.com/watch?v=8FPQTxOTW10

 

[trinibwoy]

Based on its poor yields, there isn't as much redundancy as there needed to be.

Where did you see that GK104's yields are poor?

Depends what someone means with "less efficient than GK104 for gaming". The average user won't notice much of a difference since GK110 desktop should end up on estimate about 40-50% faster than GK104.

I don't know. Given the possible arrival date far in the future and likelihood that it will hit around 50% faster than GK104 it leaves 580 owners like me in a lurch. Basically I have nothing to upgrade to for the next 7+ months. Though I'm using that reality as motivation to burn through my Steam back catalog

 

[Ailuros]

Oh you can always consider a 690 and live in AFR hell; needless to say that the grand you pay for it is a feature and not a bug

 

[Alexko]

Where did you see that GK104's yields are poor?



I don't know. Given the possible arrival date far in the future and likelihood that it will hit around 50% faster than GK104 it leaves 580 owners like me in a lurch. Basically I have nothing to upgrade to for the next 7+ months. Though I'm using that reality as motivation to burn through my Steam back catalog

Sea Islands should be here this winter.

 

[UniversalTruth]

Where did you see that GK104's yields are poor?

They have to be, that's the only reasonable explanation why there is nothing in stock.

http://www.newegg.com/Product/Produ...H&N=-1&isNodeId=1&Description=gtx+680&x=0&y=0

Depends what someone means with "less efficient than GK104 for gaming". The average user won't notice much of a difference since GK110 desktop should end up on estimate about 40-50% faster than GK104.

Well, "less efficient" means that it will offer less performance with respect to an imaginery chip which misses so crazy GPGPU capabilities.

Sea Islands should be here this winter.

Aren't they projected for 2013 too? According to some leaked images...

 

[CarstenS]

They have to be, that's the only reasonable explanation why there is nothing in stock.

http://www.newegg.com/Product/Produ...H&N=-1&isNodeId=1&Description=gtx+680&x=0&y=0

Since it's coming up again and again, I think, I'll just copy and paste what I said about availability here: It's there. You can always get one when you want one for normal prices out of reputable shops in germany and probably in other Euro-countries as well.

Just don't think that Newegg is the center of the world.

Polymorph engine isn't shown in SMX when I see block diagram.
Does it mean that GK110 will be used only for computation purpose?

Polymoprh is just an abstract for functions performed by a couple of units that are just there anyway.

 

[silent_guy]

Polymorph engine isn't shown in SMX when I see block diagram.

Does it mean that GK110 will be used only for computation purpose?

It's a compute oriented conference and a compute oriented white paper.

 

[UniversalTruth]

Since it's coming up again and again, I think, I'll just copy and paste what I said about availability here: It's there. You can always get one when you want one for normal prices out of reputable shops in germany and probably in other Euro-countries as well.

Just don't think that Newegg is the center of the world.

Actually, it is the centre... of the USA.
So, because we chew and chew, and chew one and the same thing many times.
Is there anyone from NV reading and writing here to, please, tell us why it is so hard to buy a 680 in there while it's available in Europe?

Thanks.

 

[Kaotik]

Aren't they projected for 2013 too? According to some leaked images...

If you're referring to the "platform roadmap" etc, they're not exactly accurate, for example Evergreen was "2010" despite half the series being out 2009, Northern Islands was "2011" again despite half the series launching 2010, and Southern Islands was 2012 (though that's quite accurate, as only "paper launch" happened 2011)