Kepler Nvidia GeForce GTX 780

[nr4]

Manicdan I still dont get how NVIDIA didnt hire you to show them "The Way It's Meant to be Played".
Such a big loss for them

[DilTech]

i still dont get these comments. right now that super gpu would sell for thousands in the server world. per mm2 it would be more profitable than anything else out there. but not only would it earn them lots of cash, it will let them set world records or gain crazy market share.

if they wait just 1 year the value of those super gpus drops like a rock to about where things are now. sure yeilds will be higher, but R&D wont change, and their per chip margin will be severely hurt. if they wait 2 years its almost worthless. getting something out AHEAD of schedule in the world of semiconductors is the best thing ever. delaying that costs millions.

Except GK110 DOES hit the server world in November at the absolute latest, at a time-frame where it still will have NO competition. Gave them plenty of time to get yields perfect before going any further with said chip, as well as time to work on power draw/heat.

Not only that, when you can sell a card that costs you LESS money to produce than your $200 part last year for $400-$500 this year AND move those in quantities, why would you do anything that ends that hustle?

[Lanek]

Except GK110 DOES hit the server world in November at the absolute latest, at a time-frame where it still will have NO competition. Gave them plenty of time to get yields perfect before going any further with said chip, as well as time to work on power draw/heat.

Not only that, when you can sell a card that costs you LESS money to produce than your $200 part last year for $400-$500 this year AND move those in quantities, why would you do anything that ends that hustle?

At the same time they have delay all their future chips on professional market, ( including Kepler intially )... Maxwell ( 2014 ).. November is not an hasard. I dont think Nvidia push back their prevision, just cause they can compet with lower chips in the gamer segment.... They could release thoses cards for professional, and just release GK104 for gamers...

What do you think? they was affraid gamers will buy a GK100-110 Tesla at 4000$ ?

I will not be so sure, they will not have any competition when releasing the K20 Tesla. The Tesla K20 is annunced to get a 1.5Tflops DP theorical peak at 1/3 rate. The AMD 7970 have 1 Tflops DP allready at 1/4 rate .... 1.5Tflops DP at 1/3 = 4.5Tflops SP.... And we are speaking about the GK110 .. we have no idea what was the initial GK100 who have been it seems completely lost somewhere between august to december 2011... ( If the AMD 7970 was set at 1/3 rate, they will allready hit 1.33Tflops DP ( pure theory as i absolutely dont know if AMD can do 1/3 ).

Vs Intel, the problem is even more different, the Xeon-Phi x86 is not totally comparable of GPU's...

Computing is totally something aside anyway, we can base ourself on a simple series, betwen AMD HSA who try to build something who will not take place before 2014 maybe, Intel and his Xeon-Phy who will be an entry but not the last evolution of it.. this take time and the work to do is not based on a series, but in a long process of evolution.

[Manicdan]

Manicdan I still dont get how NVIDIA didnt hire you to show them "The Way It's Meant to be Played".
Such a big loss for them

i have no idea what your saying here.
TWIMTBP was very controversial, and has nothing to do with releasing a product when its ready vs waiting until its cheaper and also less valued.

[jimba86]

I just hope this new GTX 780 is good at 3d work (maya/PS/max etc) when it comes out.

the GTX 680 sucks at this and is the reason why i dont own one (along with the fact it OC's crap due to GPU boost)

[Karolis]

i have no idea what your saying here.
TWIMTBP was very controversial, and has nothing to do with releasing a product when its ready vs waiting until its cheaper and also less valued.

What I think nr4 was trying to say is that if nvidia saw that it's financially viable to release it now, they would do it.

[STEvil]

Except GK110 DOES hit the server world in November at the absolute latest, at a time-frame where it still will have NO competition. Gave them plenty of time to get yields perfect before going any further with said chip, as well as time to work on power draw/heat.

Not only that, when you can sell a card that costs you LESS money to produce than your $200 part last year for $400-$500 this year AND move those in quantities, why would you do anything that ends that hustle?

The more time you give your competition to catch up the more money you lose due to not selling product. If you're improving the architecture then do that, and release the improved version down the road. Sitting on product does nothing but cost money.

[Olivon]

GK106-400 -> GTX 660
GK106-250 -> "GTX 660 SE": 768SPs + 192-Bit (5Gbps?)
GK106-200 -> "GTX 650 Ti": 576-768SPs + 128-Bit 5Gbps
GK106-875 -> Quadro

http://we.pcinlife.com/thread-1978270-1-1.html

Via 3DCenter, special thanks to them, really a great team

[Andrew LB]

if they wait just 1 year the value of those super gpus drops like a rock to about where things are now. sure yeilds will be higher, but R&D wont change, and their per chip margin will be severely hurt. if they wait 2 years its almost worthless. getting something out AHEAD of schedule in the world of semiconductors is the best thing ever. delaying that costs millions.

That makes absolutely no sense. Holding back the release of a super fast GPU (due to yields, no competition, etc) would have absolutely no effect on it's retail pricing. All that matters is the chips performance when compared to it's competition on release day. If a company held back release of some super-GPU a year or even two, and upon release it still destroyed the competition... pricing would still be whatever the market dictates regardless of the delay. It's not like they would be required to sell it at a price point as if it were a one or two year old product. In reality, such a delay would allow for the manufacturing process to become very good, driving the price/wafer/chip much lower, which in turn would result in the profit margin for the super-GPU to be far higher than if it were released back when it was not necessary.

[Andrew LB]

The more time you give your competition to catch up the more money you lose due to not selling product. If you're improving the architecture then do that, and release the improved version down the road. Sitting on product does nothing but cost money.

Not true. Especially when your FAB is having problems manufacturing such a complex chip at a reliable rate, causing the cost/chip to be very high. Additionally, if you already have the top performing product in the market segment with zero competition in sight, it would smart to maximize profits on that GPU before releasing a product that will cause a huge price cut to the current line as well as a reduction in sales figures.

[tajoh111]

The more time you give your competition to catch up the more money you lose due to not selling product. If you're improving the architecture then do that, and release the improved version down the road. Sitting on product does nothing but cost money.

Remember Nvidia doesn't update their professional line if you look at the past with the GX11X chip. They usually make one professional product and let it hold out that generation. Releasing GK100 in a state that is highly downclocked and disabled might do more harm than good at this point.

Releasing the GK100 would definitely devalue the rest of Nvidia's pro line. Nvidia probably has loads of quadro 6000 and 5000 products stored up considering how many Fermi products they have on shelves still. And in addition, quadro 6000 products are still highly competitive with AMD best and newest stuff(w9000). In hothardware review, the quadro takes it and in the toms hardware review, it is basically a tie. Considering that Nvidia still has the driver advantage in the pro market(this isn't questionable at all) which is huge, quadro 6000 and 5000 are still highly desirable products. A product they can sell for the top of the line price, but only have a handful of supply does more damage when they need to cut the price of their previous card in half and they half lots of supply of. If quadros based on fermi were simply uncompetitive with AMD's newest stuff, Nvidia hand would be forced much more.

One more thing that has to affect profitability is Nvidia is paying per wafer this generation and not per working chip like last generation. Releasing a low yield chip would create marginal revenue because of supply and decrease the price of everything they have. GK104 lines would likely be effected when basically every review website mentions GK100 as the top line but not being used in all but the pro line because it is too expensive to put into their consumer cards. And all of sudden the 499 dollar price tag doesn't make as much sense anymore as you get people wanting top of the line waiting instead of settling for gk104 this generation. A chip like gf100 that was low yielding last gen but still profitable because of the pay per working chip policy suddenly becomes unmanufacturable because the yield matter more than anything under this new contract at 28nm.

Considering how gk104 clocked and supply came(supply is actually decent considering the demand of 28nm wafers), I suspect a gk100 with similar flaws to gf100 could have been released maybe even better(this gen seems more energy efficient). But with the contract change with TSMC, it would have been unprofitable and did more harm than good. Releasing a polished gk110 is probably the smart move this generation as it will allow them to release a product that demolishes w9000 and compete against what Intel has lined up. When I say this, remember Nvidia has only been releasing 1 pro card based on either gx10x or gx11x and not both. They only take one shot to probably prevent pissed off customers and to justify the price tag of these cards.

[=SOC= Admiral]

Im one of those that are still waiting as you stated. But then I have a GTX 580 which is doing just fine still. And im not doing any distributed computing right now because of the cost of power where Im living. I moved back with my parents and they told me that the power bill had jumped $200. I turned off my bitcoin mining and folding@Home. But I have talked to my friend who has a 680 and he says it sucks at the Bitcoin so no reason to get one.

[tajoh111]

Im one of those that are still waiting as you stated. But then I have a GTX 580 which is doing just fine still. And im not doing any distributed computing right now because of the cost of power where Im living. I moved back with my parents and they told me that the power bill had jumped $200. I turned off my bitcoin mining and folding@Home. But I have talked to my friend who has a 680 and he says it sucks at the Bitcoin so no reason to get one.

I don't get the concept of bitcoining and it just seems like an economy waiting to collapse. The only time it seems seriously profitable in today's market is when someone else is paying the power bill. I.e Work, parents or living in a dorm and in either points it is basically stealing.

In addition because of is decentralized and not controlled by a government or country, I don't see how it will continue to have value. Major countries have currency that are worth a damn because they have resources and services to trade. Removes those resources and services and that currency is worthless as people in that country have nothing to exchange with other countries and nothing to trade with each other within their own country. Pretend a new nation pops up and says, we want to exchange some currency from your country to ours so we can buy stuff in your country. One of the first questions will be what does your country have to trade. If that country has no resources or services to offer and is simply exchanging its currency, it will be worthless.

Bitcoins just seems like an ecological waste of energy as no useful work is being done. I have seen one argument where they say bitcoing is no less harmful than farming for gold. But at least with gold, your obtaining a resource that is desired all around the world without question and is resellable for luxury or industrial applications. It is desired by the general public at large and hence its value. Your also getting a physical item when you mine it rather than a virtual one. With bitcoining the value is determined largely on demand of this virtual currency and without legs because it has no application, I can't seeing this demand being sustained.

If bitcoining was a giant network which somehow did super computing work for corporations/government and people paid for those services I could see more value in the currency as it would be doing something useful and providing a service. But as is, what benefit is it into making people happy or providing a useful service for people?

[STEvil]

Not true. Especially when your FAB is having problems manufacturing such a complex chip at a reliable rate, causing the cost/chip to be very high. Additionally, if you already have the top performing product in the market segment with zero competition in sight, it would smart to maximize profits on that GPU before releasing a product that will cause a huge price cut to the current line as well as a reduction in sales figures.

Bad FAB/R&D means burning $$ while you wait for your new part to become viable. IF it becomes viable before passing the point of return on input on R&D and you are not selling it then you are losing more money on it, let alone if you're sitting on your butt thinking your competition is having a picnic. This is why AMD had the 4870 then later the 4890. The released the 4870 because it was viable then released the updated version later as the process matured. IF the 780 is viable then nVidia should be producing and selling it asap, while applying updates to the technology to market as the 781, if it is going to be viable to do so before the next refreshes. At this time it seems the 780 has been in the shop for a really, really long time..

If the chip performs as expected and is viable then its release shouldnt hurt the current technology as its market price would likely be significantly above that of the current hardware. Unless its nV's bulldozer that is..

Tajoh has it mostly correct, its the pro line that would suffer (if anything), but really thats nV's own fault for having too much product, and honestly at the retarded prices of pro hardware, they can eat the cost.


As to bitcoining, its just a black market currency.

[Andrew LB]

Bad FAB/R&D means burning $$ while you wait for your new part to become viable.

Since when does it cost money to NOT place an order? Rushing to be the first to release big complex chips based on a very young and costly cutting edge manufacturing process is simply foolish. Both nVidia and ATi did this back when Fermi and HD5xxx were released, resulting in very limited product availibility and dismal earnings for close to half a year. Bragging rights are all fine and dandy until you have to explain to the board of directors and all your share holders why earnings are flat and why you chose to show the world your e-peen instead of doing what they hired you for.... to make money.

As I said earlier, if you already dominate the market ... it would be foolish to release a brand new costly chip ,manufactured on an immature process, where you'd essentially be competing against your own product. A smart manager would stick with the chips currently made on the very reliable and mature, yet slightly larger process where there are little to no wafer-loss... thus allowing you to maximize profits. Additionally, it gives your third party manufacturers time to sell off the any cards on store shelves or in their warehouses.

[LordEC911]

Since when does it cost money to NOT place an order?

Because there are still costs associated to that product you aren't ordering, you are wasting valuable resources that would be of better use working on the next project and, most importantly, by the time you do decide to manufacture and launch your "held" product your competitor may have something even better out or about to come out.

This is all about design cycles and deadlines. In the technology world you never, let me repeat myself, NEVER hold back a sure thing.
Sure, there are time things go bad or wrong. Yes, there might be some rare extremely short-term strategy.

You never hold back something for 6+ months because "it is too good." Look at Intel for the past 6 or so years if you want some proof.

[mdzcpa]

Sometimes you do.

When you calculate that the value of your product in the pipe will drop when the new technology is released, and that drop in value exceeds the combination of demand and price premium for the new release, you hold off. This is especially true when your current yields are high along with robust continuing retails sales and price points. You in essence have a cash cow that needs to be milked before you kill it. You want to work those large gross margins for as long as you can. If yields are not high enough on the new tech to provide wide enough margins that offset your cash cow losses you hurt yourself with the release. Not saying this is the case here, but a likely explanation and a very basic formula for product release timing.

[LordEC911]

Sometimes you do.

When you calculate that the value of your product in the pipe will drop when the new technology is released, and that drop in value exceeds the combination of demand and price premium for the new release, you hold off. This is especially true when your current yields are high along with robust continuing retails sales and price points. You in essence have a cash cow that needs to be milked before you kill it. You want to work those large gross margins for as long as you can. If yields are not high enough on the new tech to provide wide enough margins that offset your cash cow losses you hurt yourself with the release. Not saying this is the case here, but a likely explanation and a very basic formula for product release timing.

That is short-term delay and that is not what I would call a sure thing...
What you are describing is not something "being held because it is too good" but something that has a problem.

[STEvil]

Because there are still costs associated to that product you aren't ordering, you are wasting valuable resources that would be of better use working on the next project and, most importantly, by the time you do decide to manufacture and launch your "held" product your competitor may have something even better out or about to come out.

This is all about design cycles and deadlines. In the technology world you never, let me repeat myself, NEVER hold back a sure thing.
Sure, there are time things go bad or wrong. Yes, there might be some rare extremely short-term strategy.

You never hold back something for 6+ months because "it is too good." Look at Intel for the past 6 or so years if you want some proof.

I think thats the problem, they're thinking short-term market strategy. Short term is iffy at best and the big stock brokers dont like watching you play that game, especially if your "new" technology is 6 months old and hasnt even seen retail yet..

[tajoh111]

http://www.hpcwire.com/hpcwire/2012-...rcomputer.html

Looks like gk110 or gk100 is already in a usable form. Big customers are getting first dips.

32 k20 have been shipped to Oakridge with another 1000 coming next week. Seems like its manufacturable at this point, but perhaps not at the consumer level profit wise.

[Lanek]

http://www.hpcwire.com/hpcwire/2012-...rcomputer.html

Looks like gk110 or gk100 is already in a usable form. Big customers are getting first dips.

32 k20 have been shipped to Oakridge with another 1000 coming next week. Seems like its manufacturable at this point, but perhaps not at the consumer level profit wise.

The system will not be finalized before March 2013... they got thoses for tests only.. ( they need test the nodes system with AMD CPU 16cores ).. If Nvida want to release the K20 on november december, I hope they have ES samples who are running lol .. cause HPC systems need more of 6-7months of testing for be finalized. They dont build it in 1min 30sec like we do for our pc.

[R101]

Still no info about will 780 cards be cuda-neutered like 680 or not..

[LordEC911]

The system will not be finalized before March 2013... they got thoses for tests only.. ( they need test the nodes system with AMD CPU 16cores ).. If Nvida want to release the K20 on november december, I hope they have ES samples who are running lol .. cause HPC systems need more of 6-7months of testing for be finalized. They dont build it in 1min 30sec like we do for our pc.

That's what I got from that too. Early production samples, i.e. actual production silicon, to get real world data on the system and finalize software.

Still no info about will 780 cards be cuda-neutered like 680 or not..

Not sure what you mean... Consumer GTX780's will have cutdown DP but probably not to the extent of the GTX680s, probably somewhere around 1/8th to 1/16th of FP32.

Edit- Ah... I think I found what you are talking about. Nvidia didn't spend too much timing optimizing their software for compute situations and focused mainly on gaming since that is what the GTX680 is for. I'm going to go with no, GTX780 will not be CUDA neutered like GTX680.

[tajoh111]

The system will not be finalized before March 2013... they got thoses for tests only.. ( they need test the nodes system with AMD CPU 16cores ).. If Nvida want to release the K20 on november december, I hope they have ES samples who are running lol .. cause HPC systems need more of 6-7months of testing for be finalized. They dont build it in 1min 30sec like we do for our pc.

I did mention they are also getting 1000 more k20s next week. That seems like a lot of cards to just be engineering samples. The first 32 I agree are likely not final silicon. The 1000 those I think are more likely to be some sellable form of big keplar.

[Lanek]

I did mention they are also getting 1000 more k20s next week. That seems like a lot of cards to just be engineering samples. The first 32 I agree are likely not final silicon. The 1000 those I think are more likely to be some sellable form of big keplar.

Dont worry i have read the article too, but with more of 15000 GPU cores on final, 1000 is just 1/15th of the final systems. ( +18000 CPU )... thoses are early samples for test the system and be ready then to migrate the rest. Its not important they are fully functionnal or work at low speed or different speed, they just need them to be working for start the tests. This said, it is fully possible they have got the first raw of full tested chips when Nvidia is starting the production ramp up. They need a lot of thoses cores for start testing and building the system, because you cant simulate each parts as a valid results for the interconnect and the software ofc.

Yes they are some sillicon ready, as i was mention, they release the K20 in less of 2 months, i hope they have start product them lol .