Where do you get this from?
The 680 GTX has more shaders than the 670 and is only 3% faster. The 2GB b 3GB of Vram wont make much difference in anything but the biggest resolutions
No process shrink either so thats out of the question.
Basically all you get is a few more of what you got now which means about 14% better performance. No way you will get 30%
So what is everyone's msrp guess I'd gtx780 is based on gk110? I am guessing $599. I think the first lower binned part will be called gtx770 ti and will be $499. GF114's top model will be the gtx760ti and will be around $399. I think 4-6 months after gtx780 comes we'll get a higher performing (less fused off smx's, higher clocks) gk110 part, a la gtx785.
He gets it from basing the gtx780 on gk110 and not gk104/114. His guesstimations are accurate, even conservative IMO.
Where do you get this from?
The 680 GTX has more shaders than the 670 and is only 3% faster. The 2GB b 3GB of Vram wont make much difference in anything but the biggest resolutions
No process shrink either so thats out of the question.
Basically all you get is a few more of what you got now which means about 14% better performance. No way you will get 30%
Lol. You must be joking.
So a HPC GPU which is designed to run Tesla Cards is going to beat its top of the line home desktop gaming GPU? LOL
The die area is 2x as big as GK104 and the power consumption is massive. Whilst offering no tangible performance benefits in 3D gaming.
There is a very good reason why Nvidia stripped out most of the HPC from the latest desktop GPU's. What makes you think putting them back in is going to improve things for 3D gaming?
How much do you think its going to cost to make GPU's 2x as big as they are now? 2x as much thats how much.
30% is dreaming
K20X has close to 30% more computational power than the GTX680 (14/8 SMX * 0.732 GHz/1.006 GHz). Due to GK104 turbo let's say 25%. It also has 30% more bandwidth.
The 680 is up to 10% faster than the 670. If your game/setting is shader bound, it will be more, if it is bandwidth bound it will be less. You cannot give a card more compute power without bandwidth and expect it to perform significantly better - and vice versa.
So if you increase compute power and bandwidth both by 25%, you will get 25% higher performance. That is not including higher clocks for the GTX780 and possibly a full 15 SMX SKU. As I said - 30% are guaranteed for a GK110 SKU, make no mistake there.
Are you trying to baffle me with bull shit or something?
GK110 has most of its die area taken up with HPC. How will HPC help drive more frames in Crysis 3? They took that stuff OUT! in these GPU's
AMD already have 3GB 384bit Memory bandwidth and it doesnt make much difference in todays titles. A 670 GTX beats a 7950 everytime.
At best you might get 20% better FPS and its more likely to be 10-15%. Both GPU's will be 28nm and just more of the same we have now. Higher boost and a few more of everything
= complete fantasy. As Sontin said, the structure of GK110 is known and it doesn't really differ much from GK104.GK110 has most of its die area taken up with HPC.
Why not $499? It's perfect price for nVidia because AMD can't really hurt them with their line up.
High prices of GTX260 and GTX280 were the reason why AMD forced nVidia to cut their prices after 4 weeks.
Fx1,
read nVidia's pdf papers about GK110. It's has the same architecture like GK104. It's easy to predict where a Geforce product will land.
Look in the past to see the future.
I guess that helps than a discussion about GF110 and why it was earlier on the consumer than on the HPC market.
Yes apart from the fact that GK110 is 2x as big as todays GPU's upon which Nvidia has poor 28nm yields and when you make the GPU bigger there is going to be even less yield and cost double per GPU. There is no commercial sense to use a less profitable GPU. The only way they would do this is if they have a few million Tesla GK110 reject cores sitting about where parts of the GPU have been turned off.
The GPU has 7bln transistors and is a fabrication nightmare apparently. Last time Nvidia tried to do this on 40nm and they started getting 2% yields and the best they ever had was about 20%.
There is a ton of information about Nvidia on this around the web. A 780GTX will be a faulty core because they wont sell a perfect GK110 to anyone other than Tesla customers.
Its not on the consumer market. Its on the HPC market as those Tesla Cards.
Its a monster GPU that uses 350w of power. The heat will be insane and so will the costs associated.
Thats ok when your selling to enterprise. Not so good when its inside your gaming rig.
You don't know yields for Nvidias GPUs, 40nm or 28nm. I guess you are basing your claims on Charlie Demerjan from Semiaccurate who is a known Nvidia-phobe. There is simply no proof to what you are saying.
As far as my sources go, GK110@GTX780 was not profitable for launch in September 2012. It is 4 months later now and the launch might not take place for another 2-3 months. Surely in that timeframe improvements can be made.
Then:
There is no 15 SMX Tesla to begin with. The professional market uses much less cards than the gaming market. While demand for K20(X) seems to be high, it will be fulfilled sooner or later. Wafer output also increased substantially in Q3 2012 and I would expect it will increase further towards Q2 2013.
Finally:
Nvidia needs a strong GPU to compete with HD8970. They cannot do that with a GK114 or GK204 or any other kind of GK104 derivative since it has too little bandwidth. If they were to build a GK104 derivative with a 384bit memory bus, more units etc. they would basically have a GK110.
To develop and tape out a GPU costs much much money. GK110 exists, no further development costs are needed. An additional "HPC-less" GK110 would incur additional unnecessary costs AND due to the limited volume of professional solutions, GK110s R&D costs would probably never be covered. So you see, it seems to be a bad idea not to use GK110.
Please stop inventing stuff, stop lying. GF110 came in the GTX580 and the GTX570, already forgot?
Nvidia didn't release a single-GPU card in their entire history that had a TDP of 350W and they will not start now. Your claims are completely uninformed and very poorly thought out.
There was same stuff all over the net about the 685 GTX which never happened. supposed to be based on the GK110.
The only reason they might use that GPU is because they have a bunch of broken Tesla cores they cant sell. This isnt exactly a good thing for the consumer either. Neither is a 350w GPU
Who is talking about GF???? 580? who the heck mentioned those cards?
Look in the past to see the future.
I guess that helps than a discussion about GF110 and why it was earlier on the consumer than on the HPC market.
Its not on the consumer market. Its on the HPC market as those Tesla Cards.
Its a monster GPU that uses 350w of power. The heat will be insane and so will the costs associated.
Thats ok when your selling to enterprise. Not so good when its inside your gaming rig.
Who is talking about GF???? 580? who the heck mentioned those cards?
So we take rumors as reliable sources for information, now?
I have explained why they probably will use GK110, read again or ask if you didn't understand it.
There is no 350W GPU, K20 has a TDP of 225W and K20X has 235W - a far cry from your imagined 350W. Stop lying please. :thumbsdown:
Sontin was talking about GF110 and you answered. GT200, GT200b, GF100, GF110 all were released in the professional market AND in the consumer market as GeForce. That is an undeniable fact and that makes it likely that it will be the same with GK110. The only difference this generation is, that Nvidia postponed their large GPU to a point in time when it could be made with little trouble.
Well, if the GK114 has a 384 bit bus, I'll save my money for that. I want a larger monitor and the larger bus will help at larger pixel counts. Also, if GK114 is 5B xtors, then I doubt NV would put out a consumer GK110. They'll get better yields and profits off the smaller dice.
Except that any chip will have to have the appropriate cache system to work with that bandwidth effectively. Just a wider bus doesn't cut it. And whoops, we're closer to GK110 already.
Does anyone in their right mind think that Nvidia develops and tapes out GK110 for $200m (my guesstimate, anyone got info how much that roughly costs?) and then sells maybe 100k units only? That just doesn't make sense. As far as I know, the professional market is NOT enough to cover R&D cost for a large GPU on its own. That is why Nvidia released ALL their big GPUs in the consumer market too - to break even and make a profit.
Nvidia has 7000 employees. If only 1000 work(ed) on GK110 for an average salary of $60k/year for 3 years, that is 180 million dollars right there. And Keplers development began 7 years ago (of course not full force - according to Nvidia, development got really manpower intensive 3 years ago). That doesn't include utilities, internal and external expenses and process developing costs at TMSC.
Look at Intel:
They invest a shitload of money into their R&D and fabs and therefore need volume sales to make a buck. If Intel only sold Xeons, they couldn't afford their state-of-the-art fabs and their expensive research. Volume is the key word here just as with GK110.