Nvidia Debuts GK110-based 7.1 Billion Transistor Super GPU

[dennisburke]

[citation][nom]blazorthon[/nom]What do you mean by that? Kepler's prices (excluding the GTX 690) are record-breaking performance for the money, especially the GTX 670.[/citation]

Like you said in your previous post, the PCB's all seem to be tiny compared to past designs, and that the costs of production has probably been reduced. The idea of going smaller always implied that costs should go down with an improvment in efficiency, irregardless of performance. The only reason I can see for the high cost is TSMC's markup and/or Nvidia pricing according to AMD's current lineup.

 

[blazorthon]

[citation][nom]dennisburke[/nom]Like you said in your previous post, the PCB's all seem to be tiny compared to past designs, and that the costs of production has probably been reduced. The idea of going smaller always implied that costs should go down with an improvment in efficiency, irregardless of performance. The only reason I can see for the high cost is TSMC's markup and/or Nvidia pricing according to AMD's current lineup.[/citation]

Except for the 690, prices AREN'T high for their performance. You don't expect them to cut price for the performance by more than half or so every time new cards come out, do you? The GTX 670, at $400, has more performance for the price than any card that I can think of, ever, at least among the high end.

Unless you are saying well why isn't Nvidia selling these cards for like $100 for some reason, I don't see what you're saying. Nvidia has undercut AMD's price for performance with both the launch of the 680 and then again with the launch of the 670 and TSMC isn't marking anything up. The profits are going to Nvidia.

 

[dennisburke]

[citation][nom]blazorthon[/nom]Except for the 690, prices AREN'T high for their performance. You don't expect them to cut price for the performance by more than half or so every time new cards come out, do you? The GTX 670, at $400, has more performance for the price than any card that I can think of, ever, at least among the high end.Unless you are saying well why isn't Nvidia selling these cards for like $100 for some reason, I don't see what you're saying. Nvidia has undercut AMD's price for performance with both the launch of the 680 and then again with the launch of the 670 and TSMC isn't marking anything up. The profits are going to Nvidia.[/citation]

As far as TSMC's markup, one can only guess. My 570 cost me $350 and I thought that a fair price. But the whole idea of going smaller is to help reduce price, and with not as robust a PCB I can only assume other factors are leading to costs today, but I'm sure Nvidia is out to make money too.

 

[blazorthon]

[citation][nom]dennisburke[/nom]As far as TSMC's markup, one can only guess. My 570 cost me $350 and I thought that a fair price. But the whole idea of going smaller is to help reduce price, and with not as robust a PCB I can only assume other factors are leading to costs today, but I'm sure Nvidia is out to make money too.[/citation]

The 670 is about as fast as 570 SLI, yet is only $50 more than the 570 when you bought it. Nvidia is pricing their cards as they are purely because they want to make more money. Nvidia made sure that the cost of manufacturing the Kepler cards was dirt cheap so they could not only meet or beat AMD, but also continually undercut AMD in cost while keeping decent profits.

 

[MANOFKRYPTONAK]

I knew the GK-110 was going to be introduced this generation as a compute card, (at least for now, I didn't say there would never be a GK-110 gaming card).

P.S. Does this mean that Nvidia will have the graphics crown and the compute crown, or do you guys think AMD has something that'll be out by Q4 2012.

Also does this mean fiscal Q4 2012 or Q4 according to the Gregorian calendar?

 

[blazorthon]

[citation][nom]MANOFKRYPTONAK[/nom]I knew the GK-110 was going to be introduced this generation as a compute card, (at least for now, I didn't say there would never be a GK-110 gaming card).P.S. Does this mean that Nvidia will have the graphics crown and the compute crown, or do you guys think AMD has something that'll be out by Q4 2012. Also does this mean fiscal Q4 2012 or Q4 according to the Gregorian calendar?[/citation]

I think they refer to the calendar when they say Q1/2/3/4 on this site, not the fiscal quarters.

 

[mamailo]

A GK-110 gaming card would need another die shrink or two;perhaps when 14nm production nodes become reality.Also add another 500mill transistors for the render back end, compute cards dies do not include renders.
Although GK-110 can stream 8 fluxes of 4 frames @60ms (roughly 66 FPS non multiplexed).
Thats a lot of computer power.

 

[blazorthon]

[citation][nom]mamailo[/nom]A GK-110 gaming card would need another die shrink or two;perhaps when 14nm production nodes become reality.Also add another 500mill transistors for the render back end, compute cards dies do not include renders.Although GK-110 can stream 8 fluxes of 4 frames @60ms (roughly 66 FPS non multiplexed).Thats a lot of computer power.[/citation]

Why would it need another die shrink? It's not Nvidia's largest GPU to date and it probably isn't the most power-hungry. Even adding 500M transistors probably wouldn't make it the largest Nvidia GPU (although that would be close).

 

[mamailo]

[citation][nom]blazorthon[/nom]Why would it need another die shrink? It's not Nvidia's largest GPU to date and it probably isn't the most power-hungry. Even adding 500M transistors probably wouldn't make it the largest Nvidia GPU (although that would be close).[/citation]

Due the thermals.
Bigger transistors use more power even is the die areas are the same.
Today; would break tech and protection to consumer agencies specs.Server hardware is a leap frog in quality and reliability compared to consumer gear.

Is possible at hardware level but price would be very high. In current node you would need a custom power supply method; special cooling and a Framebuffer/TMDS chip for deferred rendering.

And do not forget the development/certification of the drivers

 

[blazorthon]

[citation][nom]mamailo[/nom]Due the thermals.Bigger transistors use more power even is the die areas are the same.Today; would break tech and protection to consumer agencies specs.Server hardware is a leap frog in quality and reliability compared to consumer gear. Is possible at hardware level but price would be very high. In current node you would need a custom power supply method; special cooling and a Framebuffer/TMDS chip for deferred rendering.And do not forget the development/certification of the drivers[/citation]

I'm not so sure about that. I don't think that the GK110 would be much hotter than the GF100, if hotter than it at all. Besides, if it is too hot, then all Nvidia would need to do is lower the clock frequency and/or disable a few cores like they did with the GTX 480. I'd expect that to bring down power consumption well enough for the heat generation to drop to acceptable levels and that's also assuming that it is that bad anyway.

 

[hannibal]

I'm not so sure about that. I don't think that the GK110 would be much hotter than the GF100, if hotter than it at all. Besides, if it is too hot, then all Nvidia would need to do is lower the clock frequency and/or disable a few cores like they did with the GTX 480. I'd expect that to bring down power consumption well enough for the heat generation to drop to acceptable levels and that's also assuming that it is that bad anyway.

Yep, the bigger problem will be how much there will be working chips per wafer. If there will be too few, the price will go sky high...

 

[mamailo]

[citation][nom]blazorthon[/nom]I'm not so sure about that. I don't think that the GK110 would be much hotter than the GF100, if hotter than it at all. Besides, if it is too hot, then all Nvidia would need to do is lower the clock frequency and/or disable a few cores like they did with the GTX 480. I'd expect that to bring down power consumption well enough for the heat generation to drop to acceptable levels and that's also assuming that it is that bad anyway.[/citation]

gk104 is 295 mm² and gf100 is 529 mm² so gk110 should be in between those.Therefore at the same consumption gk110 will reach a higher temp increase proportional to the square of the area difference, Basic thermodynamics.
Nevermind is a engineering problem just design a exotic cooling who stay below 42 db noise and 90 c and will be fine SOLDER e-fuses and t-fuses to pass regulations AND patch the ATX to server PSUs protocols ADD deferred render HACK the drivers TRICK ecc checks and all will be ready.
Since NVIDIA did no sell server chips to BMPs fit all in a secondary card and sell the kit for several grands.Is possible but not at "consumer products" prices like gaming cards.At most as colectors items.

Or just wait for another die shirk when all would become practical and affordable.

BTW:
I may be mistaken or been redesign but when watch the photos carefully seem like the geometry engine (PolyMorph) is missing from the GPCs,if that end to be true gaming on GK110 is mediocre at best.
Is what it is: a compute chip.

 

[blazorthon]

[citation][nom]mamailo[/nom]gk104 is 295 mm² and gf100 is 529 mm² so gk110 should be in between those.Therefore at the same consumption gk110 will reach a higher temp increase proportional to the square of the area difference, Basic thermodynamics.Nevermind is a engineering problem just design a exotic cooling who stay below 42 db noise and 90 c and will be fine SOLDER e-fuses and t-fuses to pass regulations AND patch the ATX to server PSUs protocols ADD deferred render HACK the drivers TRICK ecc checks and all will be ready.Since NVIDIA did no sell server chips to BMPs fit all in a secondary card and sell the kit for several grands.Is possible but not at "consumer products" prices like gaming cards.At most as colectors items.Or just wait for another die shirk when all would become practical and affordable.BTW:I may be mistaken or been redesign but when watch the photos carefully seem like the geometry engine (PolyMorph) is missing from the GPCs,if that end to be true gaming on GK110 is mediocre at best.Is what it is: a compute chip.[/citation]

Bringing down clock frequency linearly brings down power consumption more or less exponentially while decreasing performance only linearly at worst. Besides that, the GK110 doesn't use hot-clocking, so so long as it has a lower frequency than than double the GPU frequency of GF100, it has a lower frequency on the shader cores. Beyond that, there could be power efficiency improvements beyond the die shrink.

Furthermore, what the GK110 has in the way of gaming hardware doesn't matter right now. A gaming version could be made that lacks a lot of the compute optimized hardware, trading off compute performance for gaming performance without keeping both. The gaming version of the GK110 wouldn't need to be nearly as compute-oriented as the gaming version, so it could improve power efficiency over the compute version dramatically. It would need a revision, but not a die shrink.

 

[mamailo]

[citation][nom]blazorthon[/nom]Bringing down clock frequency linearly brings down power consumption more or less exponentially while decreasing performance only linearly at worst. Besides that, the GK110 doesn't use hot-clocking, so so long as it has a lower frequency than than double the GPU frequency of GF100, it has a lower frequency on the shader cores. Beyond that, there could be power efficiency improvements beyond the die shrink.Furthermore, what the GK110 has in the way of gaming hardware doesn't matter right now. A gaming version could be made that lacks a lot of the compute optimized hardware, trading off compute performance for gaming performance without keeping both. The gaming version of the GK110 wouldn't need to be nearly as compute-oriented as the gaming version, so it could improve power efficiency over the compute version dramatically. It would need a revision, but not a die shrink.[/citation]

It can be done? yes, Makes sense? not at all.
You can not underclock for ever; eventually each clock domain have to synchronize with the others.Even inside each domain; logic gates propagations times will set the bottom line and as with overclock; your millage will vary.

Besides once you are willing to spend a few grands for a high end card adding $300 for OEM cooling kit do not sound that bad. After all; native environment for gk110 is cooled by phase change systems that cost many;many times as much.Once again the ceiling is the economics.

As for reviewing gk110 for adapting; is stupid. A better start point would be expanding gk104 to...let' s say 7.1 billion transistors.
In oversimplistic way; just add GPCs (each add 384 cores; one raster engine and widens the card memory bus by 64bits) until you hit a barrier.
How many can be added to that beast before you must stop?. Is a question Nvidia have answer before anyone ask. Is 4.

The reason can be a thermal;geometric,manufacture;design; materials,etc.But if i have to bet; will put my money on a market one.

It would be another fermi dead end.Low yield, hot, expensive,Huge, low profit chips allover again.

Nothing comes for free in the silicon world, and the art of chip design is balancing technology with cost.The net result is profit, Nvidia botched this badly in Fermi, and did have a hard times.

So...no.It will not be a 2880 cores gaming card on 28nm.Nvidia is smarter than that.

 

[blazorthon]

[citation][nom]mamailo[/nom]It can be done? yes, Makes sense? not at all.You can not underclock for ever; eventually each clock domain have to synchronize with the others.Even inside each domain; logic gates propagations times will set the bottom line and as with overclock; your millage will vary.Besides once you are willing to spend a few grands for a high end card adding $300 for OEM cooling kit do not sound that bad. After all; native environment for gk110 is cooled by phase change systems that cost many;many times as much.Once again the ceiling is the economics.As for reviewing gk110 for adapting; is stupid. A better start point would be expanding gk104 to...let' s say 7.1 billion transistors.In oversimplistic way; just add GPCs (each add 384 cores; one raster engine and widens the card memory bus by 64bits) until you hit a barrier.How many can be added to that beast before you must stop?. Is a question Nvidia have answer before anyone ask. Is 4.The reason can be a thermal;geometric,manufacture;design; materials,etc.But if i have to bet; will put my money on a market one.It would be another fermi dead end.Low yield, hot, expensive,Huge, low profit chips allover again.Nothing comes for free in the silicon world, and the art of chip design is balancing technology with cost.The net result is profit, Nvidia botched this badly in Fermi, and did have a hard times.So...no.It will not be a 2880 cores gaming card on 28nm.Nvidia is smarter than that.[/citation]

You're over-exaggerating the issues associated a little, but yes, it probably won't happen. I never said that it would. I said that they could do it. Contrary to what you all may believe, I'm not entirely naive about this stuff. I realize the problems here and I'm simply saying that they could do it if they wanted to. Making a 7.1B (or thereabouts) transistor extension of the GK104 could also be done. Nvidia didn't only make the GK104 to have as many cores that it has because that is the maximum doable amount. They could make a much larger chip just like they did with the other generations, even if they don't make it quite as large as the previous Big GPUs.

What reason did Nvidia have for making a larger chip this generation for gaming purposes? None. The GK104 in the GTX 670 and the GTX 680 is proven to be adequate for 2560x1600 gaming in most situations and what point would there be for making a faster GPU? Something like the GK110 (a gaming version of it) would be able to be a part of a card that could probably do some 4K resolutions in gaming decently. Triple 1080p and 1920x1200 wouldn't even be a problem. There's not much market for a card that fast right now, let alone a single GPU card of such performance. It wouldn't just have a higher cost of manufacturing and a price reflecting that, it would have few buyers too. Whether or not Nvidia can do it isn't something that I think needs to be questioned (I have no doubt about their capability of making such a chip and making it more or less practical), but why they would do it. They had GK104 which performs very well (despite the memory bottle-necks) and is very cheap to make (as is the rest of the cards), so why make something that is faster when there would be no way of selling it in high volume?

However, Nvidia could underclock it compared to the professional GK110 if thermal output is too great (they shouldn't need to underclock it so far that your listed possible problems would become actual problems). There's no reason that they can't drop the frequency and voltage until standard cooling would be adequate if they wanted to. Whether or not they would want to do all of that for a chip to go into a card that would hardly sell at all is what keeps them from doing it. They would probably pay more making it than they could get back regardless of prices just because not enough people would buy it.

 

[snake8989]

The same length of GTX 580

 

[BestJinjo]

@ blazorthon,

NV can release a 1Ghz GK110 part with 12-13 SMX clusters OR a downclocked 2880 SP GK110, but they cannot do both at the same time. Otherwise, you end up with a 550mm^2 1Ghz part that will use 250-275W of power. I doubt they want to see a repeat of Fermi. GTX680 already uses about 180-185W of power under gaming. You cannot almost double the size of the chip on the same 28nm node and maintain the same 1Ghz clocks, while adding GPGPU/double precision/dynamic scheduler functionality and only grow your power consumption 70W. It's just not possible.

Even when you look at GTX660Ti vs. GTX680, by chopping off 25% memory bandwidth and ROPs, and still reducing the shader count 14% from 1536, it dropped power consumption 40W from the 680:
http://www.techpowerup.com/reviews/Palit/GeForce_GTX_660_Ti_Jet_Stream/26.html

Now think about a chip with 1Ghz clocks, full GPGPU capability, 48 ROPs, 2880 SPs, 240 TMUs, and 384-bit memory bus? It's going to be using 275-300W of power unless 28nm node has some trick up its sleeve.

 

[blazorthon]

[citation][nom]BestJinjo[/nom]@ blazorthon,NV can release a 1Ghz GK110 part with 12-13 SMX clusters OR a downclocked 2880 SP GK110, but they cannot do both at the same time. Otherwise, you end up with a 550mm^2 1Ghz part that will use 250-275W of power. I doubt they want to see a repeat of Fermi. GTX680 already uses about 180-185W of power under gaming. You cannot almost double the size of the chip on the same 28nm node and maintain the same 1Ghz clocks, while adding GPGPU/double precision/dynamic scheduler functionality and only grow your power consumption 70W. It's just not possible. Even when you look at GTX660Ti vs. GTX680, by chopping off 25% memory bandwidth and ROPs, and still reducing the shader count 14% from 1536, it dropped power consumption 40W from the 680:http://www.techpowerup.com/reviews [...] am/26.htmlNow think about a chip with 1Ghz clocks, full GPGPU capability, 48 ROPs, 2880 SPs, 240 TMUs, and 384-bit memory bus? It's going to be using 275-300W of power unless 28nm node has some trick up its sleeve.[/citation]

Your 660 TI versus 680 example is extremely flawed. The 670 performs about on-par with the 680 and it uses far less power than the 680 despite having the exact same memory configuration. The 660 TI and the 670, in your link, use very similar amounts of power (with the 670 using a little more). The power usage drop isn't even from reducing the VRAM interface, especially considering that the 660 Ti and the 670 both have eight RAM ICs. The power consumption drop from the 680 to the 660 TI is almost purely in the GPU. You also failed to realize that you're counting peak power consumption aka power consumption under stress tests, not gaming power consumption which is much lower on the 680 than it is at peak whereas the 660 TI doesn't change nearly as much between them. Average power consumption during gaming, the 680 would only use about 20-25% more power. That's almost exactly the difference in GPU performance.

The *685* would use more power. That is without a doubt. However, it would not need to use as much as you claim. If Nvidia doesn't want a repeat of Fermi, then the answer is obvious. Make the *685* have the combination of performance and power consumption that they want it to have. They could make a card between the 680 and the 690 in power consumption and performance if they want to and a *Big Kepler* chip would be the way to do it if they don't want a dual-GPU GK106 card.

Also, if Nvidia wanted to, they could make a chip with fewer SMXs and a lower frequency than the GK110. They could do either one or both if desired. Which one isn't my choice to make, so I made sure that the choices are on the table. Nvidia would make that choice if they make a faster single-GPU card than the 680 within the GTX 600 series.

However, like I said before, I doubt that they'd do this. The market for it simply isn't thriving. My point is that they can do it, not will do this.

 

[s997863]

But can it play Outcast??

what? you don't understand??

ok: Crysis aside, modern Geforce cards suck in backwards compatibility so much that my latest Nvidia panel can't offer the most basic functions that older 90s cards did: desktop color profiles, stretch-to-fullscreen for flatpanel monitors, especially for legacy resolutions lower than 640x480. Vsync simply doesn't work on older apps, games, video/movie players no matter what settings you pick, and the new cards make these classics buggy unusable without some 3rd party hack: X-COM, baldur's-gate and anything on infinity engine, system-shock-2 and anything on theif engine, any voxel game under 640x480. kinda like the moron developers who go on about 3d sound but forget to include simple things like left-right speaker switching option.

 

[magikherbs]

[citation][nom]fb39ca4[/nom]If it will help oil companies, its a bad thing.[/citation]

I agree.

The true potential of technology IS wasted on finding more ways to destroy the planet and our humanity.

Then there is the gaming/ terror sim industry and how they have used graphically violent games as a training ground for the future's occupying forces. If we can't find our own, we'll invade or destablize and steal theirs.

None of this will matter if the claims of The Keshe Foundation are proven to be true.

http://www.keshefoundation.org/en/

pEACe

 

[Guest]

Hurry up and release it already but don't make it 500-1000 dollars.

 

[hannibal]

Hurry up and release it already but don't make it 500-1000 dollars.

No it will be 1000-1500$ ;-)
We don't know, it depends on how fast GPU AMD can produce. If they can compete, this can be 600-1000$ if AMD can not... 1000$-2000$ is guite possible... If they can sell cheaper to produce GPUs at 600$ aka 680, they can ask these much, much more...
Pity but true... Only competition can reduce those prices, just like 650 has done to 7850 series lately!

 

[BestJinjo]

blazorthon,

I never said that NV won't launch a GK110 but I specifically mentioned it will NOT be a 1Ghz clocks, full GPGPU capability, 48 ROPs, 2880 SPs, 240 TMUs, and 384-bit memory bus part.

The Tesla K20 part is rumored to be only 2496 SPs which gives up a good idea of what we may see as a consumer GK110 GPU next year from NV. This is exactly what I said earlier - a 12-13 SMX cluster GK110 is doable at 1Ghz but NOT a 15 SMX 2880 SP part:
http://www.cadnetwork.de/konfigurator/gpu_rackserver_proviz_g13/system=88

Based on these specs, we see that it has 3.52 Tflops of single precision performance with 2,496 CUDA cores. That approximates about 700-705 mhz for the GPU clock. Obviously NV is seriously constrained by the TDP as I said earlier. I will stand firm by my comments that NV cannot physically launch a 1Ghz 2880 SP, 240 TMU, 384-bit GK110 in 2013 without going back to Fermi levels of power consumption at 250W average in games.