GTX 780 rumors

[f1sherman]

There hasn't been a single exclusively Tesla/Quadro chip so far. Ever.
They need volume to absorb R&D, so GF is almost mandatory as boxleitnerb hints.

Also, it has always been the other way around, salvaged parts were used for pro market, full for GF, with full chips coming to Pro later if ever. I think.

Then again there's a first time for everything.

@Fx1 stop bickering and read what the man says> he knows.. stuff

The only calculation you need to worry about is Price per GPU. A 7bln transistor GPU costs BIG. Thats why 1 Tesla is $3200 EACH!

So why are they going to put them in a $500 desktop card?

because you are cheap and won't give them $3200, but you might $500?

 

[boxleitnerb]

And you think developing this GPU costs nothing?
Do you think, angels descend from heaven and give Nvidia the blueprints?
Do you think Nvidia doesn't have to employ people to give support, write drivers and software, cultivate developer relationships with game makers, hollywood studios, universities etc.?

Simple calculation:
100,000 x $1500 (?) profit/unit from K20 and K20X = 150 million dollars profit
Get's eaten up by R&D cost 100%. No profit remaining.

Sell the same GPU as GeForce too with $60 profit/unit:
2,000,000 x $100 = 120 million dollar profit

So why put GK110 in a desktop card? Because K20 and K20X numbers are too low to break even or to make significant enough profit to drive R&D for the next generations. Easy as that.

Btw as for BOM of last generation graphics cards on a mature 40nm process (Q3 2011):

Source: Mercury Research

AIBs had to pay Nvidia $120 for 520mm2 GF110. Hm...not exactly terribly expensive, is it? Granted, 28nm is more expensive, but even if you slap 50% on it, you'll end up with $180 - still doable in a $549-599 card with decent margin.

 

[sontin]

I give Fx1 an award for making the most consecutive uninformed, blatantly wrong posts in a row.

Back on topic, if GK110 doesn't come to the Geforce lineup then quite simply there has to be a Kepler chip between GK104 and GK110 that has not been released yet (GK112???) But I seriously doubt that.

If they design another chip i guess it would be a GK106 x 2:
5 Rasterizer,
10 SMX - 10 geometry units,
384bit,
768 kbyte L2 Cache

Die size around 400mm^2 and average 40% faster than GTX680.

 

[notty22]

There were also Tesla and Workstation Fermi's that were basically GTX 580's/ 570's that cost 3-$5000 dollars also (ECC ram). It's those PRO prices that allow Nvidia to produce a larger die that ends up in the Gaming sector. It's called a
Business Model

 

[Fx1]

There were also Tesla and Workstation Fermi's that were basically GTX 580's/ 570's that cost 3-$5000 dollars also (ECC ram). It's those PRO prices that allow Nvidia to produce a larger die that ends up in the Gaming sector. It's called a
Business Model

Maybe your right. Maybe your not.

7bln transistors is seriously more than the 580 GTX's 3bln Fermi and 680's 3.5bln

Nvidia cant get good yields according to good reports so they have a monumental task to produce a whole generations of GFX cards based on a 7bln transistor GPU.

 

[f1sherman]

Char-lie, that you?

NV making $500+M on broken and unyielding chips last year alone
Any comment on that?

 

[sontin]

Yap:
With good yields their gross margins would be higher than Intels.

 

[f1sherman]

True that

Still kinda far from broken...

 

[notty22]

Maybe your right. Maybe your not.

7bln transistors is seriously more than the 580 GTX's 3bln Fermi and 680's 3.5bln

Nvidia cant get good yields according to good reports so they have a monumental task to produce a whole generations of GFX cards based on a 7bln transistor GPU.

And Fermi was a lot more transistors than anything before it. You are just speculating with cynicism , and not looking at past history/ which is fact not maybe's.

 

[bunnyfubbles]

The only calculation you need to worry about is Price per GPU. A 7bln transistor GPU costs BIG. Thats why 1 Tesla is $3200 EACH!

So why are they going to put them in a $500 desktop card? When they can put a GPU half its size and probably a fraction of the price on the card?

Profit margins sure are better when you sell something at $3200 vs. $500. But why stop there? Why not just price 1 GPU @ one hundred billion dollars. Then you only have to sell just one and you're set.

 

[tviceman]

I dont know why but you seem to want to back up your views with irrelevant calculations that are just plucked from obscurity.

The only calculation you need to worry about is Price per GPU. A 7bln transistor GPU costs BIG. Thats why 1 Tesla is $3200 EACH!

Chip costs are not based on transistor count, they are based on size. GK110 is about the same size as GF110.

So why are they going to put them in a $500 desktop card? When they can put a GPU half its size and probably a fraction of the price on the card?

You do realize that GF100 and GF110 were packaged as Tesla cards that sold for thousands of dollars, right? This generation is absolutely no different than the last, other than the demand for HPC Kepler cards has exceeded the entire sales of history of all previous Tesla cards combined.

The only thing we know for sure is that yields on 7bln transistor GPU's are going to be terrible.

Nvidia's gross margins have risen each quarter since Kepler first came out. If any part of Kepler was yielding bad, gross margins would not continue to rise. It's as simple as that.

You have no reliable information either so please dont call me out for looking to internet rumours.

You have no reliable information to back up your claim. In fact, based on historical actions Nvidia has *never* *ever* developed an exclusive HPC chip. So it's a much more likely scenario that GK110 won't be exclusive to the HPC market.

Fact is that a GK110 has an MSRP of $3200 which makes huge GPU's possible.

Again, GF110 sold in Tesla cards for thousands of dollars per card but that did not stop GF110 from being a Geforce.

So if Nvidia use broken GK110's as desktop GPU's then they will not actually be GK110 specced GPU's any more. Those chips will be lesser GPU's of unknown specification. So what exactly are you getting?

Do you realize that the gtx480 was not a fully functional chip? Do you realize that current K20 based GK110's are not fully functional chips? By your very definition, K20's are "broken" GPU's and aren't actually spec'd GK110's.

I still think your more likely to see a different GPU slightly bigger and faster than the current GK104.

That would be desirable, but there has been no heavily indicated rumors that allude to this. There were some fleeting rumors about it months ago, but most were discarded as being false. The fact is, if such a chip existed, it's almost guaranteed the community would know of it's existence.

Maybe your right. Maybe your not.

7bln transistors is seriously more than the 580 GTX's 3bln Fermi and 680's 3.5bln

Nvidia cant get good yields according to good reports so they have a monumental task to produce a whole generations of GFX cards based on a 7bln transistor GPU.

GF100/GF110's 3 billion transistors was "seriously more" than GT200b's (gtx285) 1.4 billion transistors. GF100/GF110 still became a Geforce card.


You really need to look over your facts before you start making uninformed assertions.

 

[Haserath]

Chip costs are not based on transistor count, they are based on size. GK110 is about the same size as GF110.

You sure?

 

[tviceman]

You sure?

It hasn't been confirmed by reliable sources, but rumors put it at ~550mm^2. GF110 was ~520-530 mm^2.

 

[boxleitnerb]

Absolute reticle limit at TMSC is around 600mm^2. I would imagine no one would go too close to that limit.

 

[skipsneeky2]

The proposed specs look pretty sexy if true,doubt even a 40% performance increase will come of the refresh with all the increment performance increases these last few generations have given us.

Maxwell with all that computing power really sounds like the thing people are waiting for,i know i am.

 

[Haserath]

It hasn't been confirmed by reliable sources, but rumors put it at ~550mm^2. GF110 was ~520-530 mm^2.

Sorry, I meant this part:

Chip costs are not based on transistor count, they are based on size.

 

[Grooveriding]

If they do use GK110 for a GTX780, I expect it will go for $600, maybe $650. They'll price against the competition. I don't expect we'll get it for $500 and I don't think 8970 would be as fast or faster than a GK110 Geforce card. So it will cost $600+

I still have my doubts they will use it for a Geforce card though after selling a 300mm2 die for $500 and doing it successfully. They could just as easily be cooking up a slightly bigger die with the usual marginal performance increase you get out of a refresh on the same node. Just because there is a GK110 doesn't mean they have to use it for geforce and unknown chips showing up out of the blue a month or so before release happens all the time.

Would be nice to get a GK110 though for geforce. The 680 didn't deliver a flagship performance increase over the 580, while GK110 no doubt would.

 

[boxleitnerb]

Sorry, I meant this part:

Chip costs are not based on transistor count, they are based on size.

The manufacturing costs are usually charged per wafer and are independent from wafer content. How many chips the buyer gets from that wafer is their business alone. Nvidia pays $6000-8000 per 28nm wafer approximately and it doesn't concern TMSC if there are ~100 GK110 or ~200 GK104 dies on it.

http://www.silicon-edge.co.uk/j/index.php?option=com_content&view=article&id=68

 

[Fx1]

If this is going to be a repeat of the 480 GTX and 460 GTX scenario where the big GPU was 400 GBP and then the 460 GTX was 170 GBP then there will be no point in buying the 780 GTX. The 2 smaller GPU's in SLI will outperform the 1 big one by a serious margin just like when the 460 GTX Sli would walk all over the 480 GTX and you spend only 3/4 of the money

Those big ass GPU cards were never that good. They run hot and clock poorly too. Usually followed by a big price tag also.

There is still a problem with yield on 7bln transistors. The more you try and pack into those dies then the more margin for error there is. This is probably why we never got the GTX 685 that was supposed to launch.

There are some tech sites who went into Nvidias earning calls who started commenting changes in the way they pay for wafers and associated costs of each GPU.

Also you can forget the 790 GTX if this is the case

 

[Fx1]

The manufacturing costs are usually charged per wafer and are independent from wafer content. How many chips the buyer gets from that wafer is their business alone. Nvidia pays $6000-8000 per 28nm wafer approximately and it doesn't concern TMSC if there are ~100 GK110 or ~200 GK104 dies on it.

http://www.silicon-edge.co.uk/j/index.php?option=com_content&view=article&id=68

This is the problem with a big GPU.

More transistors the more possible broken blocks and less up to spec GPU's per wafer.

If the GPU is 2x the size then i bet its not 2x as expensive to make i bet is more like 10x because of the diminishing returns.

 

[boxleitnerb]

This is the problem with a big GPU.

More transistors the more possible broken blocks and less up to spec GPU's per wafer.

If the GPU is 2x the size then i bet its not 2x as expensive to make i bet is more like 10x because of the diminishing returns.

Look at the BOM (bills of materials) chart I posted. On a mature 40nm process, a 360mm2 GPU (GF114) cost $50, a 520mm2 GPU (GF110) cost $120. That's not even factor 3. Let's say for Kepler the factor is 3-4 since GK104 is smaller than 360mm2 and GK110 should be slightly bigger than 550mm2. Still a far cry from your assumed 10x.

It is self explanatory that yields on a large GPU will always be worse than on a smaller GPU. That doesn't mean it isn't economically feasible to launch such a large GPU.

 

[Haserath]

The manufacturing costs are usually charged per wafer and are independent from wafer content. How many chips the buyer gets from that wafer is their business alone. Nvidia pays $6000-8000 per 28nm wafer approximately and it doesn't concern TMSC if there are ~100 GK110 or ~200 GK104 dies on it.

http://www.silicon-edge.co.uk/j/index.php?option=com_content&view=article&id=68

But it would seem like the more structures you need to work, the less likely they will all work. They could probably squeeze more in the same die size, but have it more often fail evaluation.

 

[Fx1]

Look at the BOM (bills of materials) chart I posted. On a mature 40nm process, a 360mm2 GPU (GF114) cost $50, a 520mm2 GPU (GF110) cost $120. That's not even factor 3. Let's say for Kepler the factor is 3-4 since GK104 is smaller than 360mm2 and GK110 should be slightly bigger than 550mm2. Still a far cry from your assumed 10x.

It is self explanatory that yields on a large GPU will always be worse than on a smaller GPU. That doesn't mean it isn't economically feasible to launch such a large GPU.

This is true but we are talking about 28nm which i expect is much harder than 40nm. The more transistors you add i would expect the failure rate to go up as you reduce the process node.

A short story will have less spelling mistakes than a novel. Same holds true here. Also you have to be careful with claimed cost of goods on a balance sheet, company's massage the figures to keep their share prices high.

 

[boxleitnerb]

But it would seem like the more structures you need to work, the less likely they will all work. They could probably squeeze more in the same die size, but have it more often fail evaluation.

That's absolutely true, at least in my understanding. So indirectly, transistor count can influence the cost, but we have to be careful here. Is there redundancy, can we just clock lower and make it still work etc.

As I pointed out with the BOM table, the ratio between transistor count and die size is not linear to cost, but it is far better then some people here claim. I would expect that with yields taken into account a GK110 GPU could cost $150-200. While that sounds like much, the highest end GK110 SKU will probably sell for $599 - $100 more than the GTX580. So the increased cost of the GPU could be passed on to the customer, to us. If it is fast enough, people will buy. If they spend $499 for GTX580+35% (the GTX680), they will surely spend $599 or even $649 for GTX580+90%.

 

[Ajay]

The manufacturing costs are usually charged per wafer and are independent from wafer content. How many chips the buyer gets from that wafer is their business alone. Nvidia pays $6000-8000 per 28nm wafer approximately and it doesn't concern TMSC if there are ~100 GK110 or ~200 GK104 dies on it.

http://www.silicon-edge.co.uk/j/index.php?option=com_content&view=article&id=68

^This.

Basically, assuming the same xtor density you wind up with ~x defects per wafer - so you 'waste' less wafer area with smaller GPUs. Because of that yields will a big factor in NV's business decision on whether or not to use GK110 in Geforce GPUs. The decision has all most certainly has already been made, we are just waiting around hoping for more solid info to plan around.