Nvidia unveils the GeForce GTX 780 Ti

[iMacmatician]

So now lets look at GK110 to GK180 again for the professional class of cards.

732mhz -->900mhz
2688 shaders --> 2880 shaders
235watt TDP --> 225 watt tdp

Now lets fill in the blanks for titan to gtx 780 ti
876mhz --> ????
2688 shaders --> 2880 shaders
250watt TDP --> ???

Doesn't the TITAN have a 837 MHz core clock? Even so, we get 1029 MHz core [5.93 TFLOPS] and 239 W TDP [24.8 GFLOPS/W] if the increase is proportional (probably won't be the case) and 933 MHz core [5.4 TFLOPS] and 245 W TDP [22.0 GFLOPS/W] if the increase is half that of the K20X to K6000 transition.

We can also look at memory speed. The M2070 has 3.132 Gbps memory and the 6000 has 3 Gbps memory, so somewhat close.

K20X to K6000: 5.2 Gbps to 6.0 Gbps

TITAN to 780 Ti (taking into account the difference between the M2070 and 6000): 6.0 Gbps to 7.2 Gbps [347 GB/s] (yeah right) proportional and 6.6 Gbps [317 GB/s] with half the increase.

[tajoh111]

Doesn't the TITAN have a 837 MHz core clock? Even so, we get 1029 MHz core [5.93 TFLOPS] and 239 W TDP [24.8 GFLOPS/W] if the increase is proportional (probably won't be the case) and 933 MHz core [5.4 TFLOPS] and 245 W TDP [22.0 GFLOPS/W] if the increase is half that of the K20X to K6000 transition.

We can also look at memory speed. The M2070 has 3.132 Gbps memory and the 6000 has 3 Gbps memory, so somewhat close.

K20X to K6000: 5.2 Gbps to 6.0 Gbps

TITAN to 780 Ti (taking into account the difference between the M2070 and 6000): 6.0 Gbps to 7.2 Gbps [347 GB/s] (yeah right) proportional and 6.6 Gbps [317 GB/s] with half the increase.

It has a core clock of 837 but never runs at that speed, in gaming scenarios, is typically greater than 900mhz. Nonetheless, the 732mhz clock of k20x is constant. So in practice, there is an even larger difference between titan and k20x.

There's also one more things that going to result in power savings for gtx 780 ti. Versus the k6000, it only has 3gb of memory to address rather than 12gb. They is going to likely be balanced out by using 7ghz samsung chips, rather than 6ghz hynix chips(if these choose to use difference memory).

As I said early, I am not assuming the best case scenario where the increase is proportional, but even if just a little bit, like half the increase between k6000 and k20x, we still have very impressive increase for a respin of gk110. But this depends on a couple things, they actually use GK180 and the variability between bins on gk180.

But knowing Nvidia as of late, they are going to clock it low, so they can brag about the fantastic performance per watt. But card partners can pull the slack here. With cards like a lightning gtx 780 ti or classified, I would expect cards with 1000-1050mhz clocks if only the process improvement allows gtx 780 ti to clock as well as gtx 780's. This is a pretty low bar and I would expect them to at least match this.

[iMacmatician]

But knowing Nvidia as of late, they are going to clock it low, so they can brag about the fantastic performance per watt.

I'm half-expecting the 780 Ti to have "only" a 225 W TDP.

[Dimitriman]

Ok, so where are the facts confirming that 780 Ti is going to be faster than the Titan? Is there any other evidence out there besides Oj101 says so? (no offense)

Strange how it suddenly has become fact in this thread... let me know if I missed something.

[tajoh111]

Ok, so where are the facts confirming that 780 Ti is going to be faster than the Titan? Is there any other evidence out there besides Oj101 says so? (no offense)

Strange how it suddenly has become fact in this thread... let me know if I missed something.

Jens Huang did say it was their fastest graphics card we ever built when he unveiled it. So I guess you missed that.

[KiSUAN]

Jens Huang did say it was their fastest graphics card we ever built when he unveiled it. So I guess you missed that.

And the prophet said:

Take some screws, some glue and make the most powerful video cardzzzz...



All the fanboyz hail the prophet


Guess he didn't missed anything

[Dimitriman]

That kinda sucks if you're a Titan user then. Buys a card last month for 1000 bucks and now someone else buys something faster for 650. I think if this mystic card is faster than Titan I find it hard to believe it will sell only for 650.

[tajoh111]

That kinda sucks if you're a Titan user then. Buys a card last month for 1000 bucks and now someone else buys something faster for 650. I think if this mystic card is faster than Titan I find it hard to believe it will sell only for 650.

It really does. Titans $1000 dollar pricing was ridiculous and I personally hoped it would have been a failure for sales because people buying a 1000 dollar graphics card sends a bad message to Nvidia. That being, people will buy a $1000 if no competition is present. I would hope people who were looking for that type of performance after the gtx 780 released just bought an overclocked gtx 780 rather than a Titan.


Nvidia was really conservative when they released Titan and gtx 780.

The gtx 780 had 20% of its shaders disabled which is ridiculous for a part 650 dollars. Its not even close to a full chip and its being charged as such. This is even more cut down than the gtx 470 was which was pretty cut down.

For a chip the size of GK110, its easy to make a card faster than titan when you think about it. Titan is clocked like a lame duck and as a result, there are gtx 780's which are faster than it, even with the big shader disadvantage.

Cards like the gtx 780 ACX that are 5% faster than titan at high resolutions.

http://www.techpowerup.com/reviews/E...Cooler/26.html

All at the slightly less than mythical price of $660.

As long as Nvidia uses gk110 with a decent amount of shaders enabled and doesn't clock it as low as Titan, then they should have a part that is easily faster.

And at this point what do you think is smarter?

Continue to charge 1000 dollars when your competitor has something competitive with it at 650 dollars and thus sales of that 1000 dollar product and your current 650 dollar product won't sell at all.

Or release something thats faster at the 650 dollar range to take on that competitors product and thus, get sales.

The gtx 580 was released prior to the 6970 and was 14 % faster than a gtx $480, it was faster and yet the cost was the same. Sure it might have pissed off gtx 480 owners that their card was being replaced with a card the same price, faster and better everywhere(sound and heat and overclocking). But Nvidia released it anyways 8 or 9 months later(the same as this launch).

Personally I still think the gtx 780 Ti will have less shaders than 2880. There's a good chance it won't have 7ghz memory either but what it should be is faster than Titan. There are too many OC versions of the gtx 780 that are faster than titan already at $650. Releasing something slower than titan at 650 dollars, is a regression from the current gtx 780 OC edition lineup. It won't have any impact and won't be a counter.

Well atleast those Titan owners still have 6gb of ram and non-crippled Double precision(which is the only reason they should have bought it in the first place).

[boxleitnerb]

If we examine the specs between gk110 based k20x tesla, the clocks of the cards were only 732 mhz, yet it had a rated tdp of 235watts. Pretty low clocks, but GK110 based titan still goes to 876mhz(even more in practice).

Look at GK180 based K6000 and you have a card that is clocked at 900mhz, has more shaders and has a tdp 10 watts lower TDP than k20x. That's a massive improvement. Before you say they aren't comparable since Tesla is different than the workstation quadro, lets look at the fermi generation.

Tesla M2070 and Quadro 6000 are both based on the same GF100 and if you look at the clocks, they are exactly the same at 574mhz and not only that, they have the same TDP at 225 Watts. They have the same clocks and TDP rating

So now lets look at GK110 to GK180 again for the professional class of cards.

732mhz -->900mhz
2688 shaders --> 2880 shaders
235watt TDP --> 225 watt tdp

Now lets fill in the blanks for titan to gtx 780 ti
876mhz --> ????
2688 shaders --> 2880 shaders
250watt TDP --> ???

So if any of GK180's improvements carry over, even just a little bit(and they used GK180 instead of Gk110), then you have room for monster clocked cards, just as much as gtx 780 if not more.

The 900mhz clocks and 225 watt TDP on a fully enabled big keplar is very impressive for quadro.

One important question:
Does Quadro have full DP rate or not? Afaik, DP is more power intensive, hence the lower clocks on the Teslas and on Titan once you enable DP. It was always my understanding that Quadro is a workstation card. Geometry, graphics etc. And Tesla is for DP. That's the whole sense of Maximus: Put one Quadro and one Tesla together, each with their specialized tasks.

Second question:
Will a potential 15 SMX-GeForce get GK180 or GK110 and if there really is an improved energy efficiency, is it due to design or due to better binning/process improvements over time?

Third question:
TDP != power consumption. Or better: Is TDP (Quadro/Tesla) comparable to TDP (GeForce)?

Under sustained gaming load, Titan and the GTX 780 often clock near the base clock due to the low temperature target:
http://ht4u.net/reviews/2013/nvidia_...st/index10.php
http://ht4u.net/reviews/2013/nvidia_...iew/index9.php

http://www.hardware.fr/articles/887-...ost-tests.html
http://www.hardware.fr/articles/894-...ost-tests.html

https://www.computerbase.de/artikel/...-gtx-titan/19/
https://www.computerbase.de/artikel/...80-im-test/11/

Now with those lower clocks, power efficiency is much better already. The Titan uses about 206W on average, the 780 uses 189W (cards only, no efficiency losses at the power supply, direct measurement):
http://www.3dcenter.org/artikel/eine...rafikkarten-st

I simply don't see that much potential for improving energy efficiency between comparable operating points (either no boost vs no boost or full boost vs full boost). Voltage is key here. The difference between no boost and boost is a whopping 0.16V! With my Titan, I measure 50-70W difference between base clock@1.0V and 1006 MHz@1.162V (whole system), and this is in line with the measurements of other reviews that made this kind of investigation.

[kromosto]

if 780 ti > titan and 780 ti approx 650 then this shows me how nvidia paniced because of 290x. i still don't get it totally (console deals explain some panic of nvidia but not all the panic after 290x) but of course they saw something i don't see, i am not able to see.

EDIT: maybe they didn't expected 290x to take this much attention and didn't expected amd to make this much performance card with this price.

What makes me happy is all my predictions before even release of Titan seems to come real. Of course still way to go but they are still on track. After today i will not call them predictions, i will call them forecast

[paulbagz]

Oh well, buyers remorse for buying top tier gear I suppose.

-PB

[LordEC911]

if 780 ti > titan and 780 ti approx 650 then this shows me how nvidia paniced because of 290x. i still don't get it totally (console deals explain some panic of nvidia but not all the panic after 290x) but of course they saw something i don't see, i am not able to see.

I wouldn't call it panic, more like a natural response.

[DarthBeavis]

Still installing dual Titans in my LAN rig Four 780s already in my main one. Three extra 780s perhaps to upgrade my son's 3 680s

[kromosto]

I wouldn't call it panic, more like a natural response.

if they made all these after the release of 290x yes i would call this a natural response but a fast organized event before amd's event announcing a new card with even haven't decided the price and some other things... these are all seem panic to me.

[tajoh111]

One important question:
Does Quadro have full DP rate or not? Afaik, DP is more power intensive, hence the lower clocks on the Teslas and on Titan once you enable DP. It was always my understanding that Quadro is a workstation card. Geometry, graphics etc. And Tesla is for DP. That's the whole sense of Maximus: Put one Quadro and one Tesla together, each with their specialized tasks.

Second question:
Will a potential 15 SMX-GeForce get GK180 or GK110 and if there really is an improved energy efficiency, is it due to design or due to better binning/process improvements over time?

Third question:
TDP != power consumption. Or better: Is TDP (Quadro/Tesla) comparable to TDP (GeForce)?

Under sustained gaming load, Titan and the GTX 780 often clock near the base clock due to the low temperature target:
http://ht4u.net/reviews/2013/nvidia_...st/index10.php
http://ht4u.net/reviews/2013/nvidia_...iew/index9.php

http://www.hardware.fr/articles/887-...ost-tests.html
http://www.hardware.fr/articles/894-...ost-tests.html

https://www.computerbase.de/artikel/...-gtx-titan/19/
https://www.computerbase.de/artikel/...80-im-test/11/

Now with those lower clocks, power efficiency is much better already. The Titan uses about 206W on average, the 780 uses 189W (cards only, no efficiency losses at the power supply, direct measurement):
http://www.3dcenter.org/artikel/eine...rafikkarten-st

I simply don't see that much potential for improving energy efficiency between comparable operating points (either no boost vs no boost or full boost vs full boost). Voltage is key here. The difference between no boost and boost is a whopping 0.16V! With my Titan, I measure 50-70W difference between base clock@1.0V and 1006 MHz@1.162V (whole system), and this is in line with the measurements of other reviews that made this kind of investigation.

Quadro has uncrippled double precision like k20x. Its was one of the main bullet points for it. FP64 is set at 1/3 like the rest of the keplar Tesla line up, not 1/24 like the gaming cards.

http://www.anandtech.com/show/7166/n...s-quadro-k6000

That is what makes the frequency and TDP so impressive. Although TDP doesn't exactly mean power consumption there is a correlation and heat(which is TDP) was one of the things sort of holding back clocks. So improved TDP = improved clocks.

No one knows the exact specs of gtx 780 ti, just that its going to be faster.

But there is one thing that should tell us if it is based on gk110 or gk180, the number of enabled shaders. My impression was gk110, it was difficult to say the least to get 2880 shaders. Thus for gtx 780 ti to be mass producible and it has 2880 shaders, it has to made on a respin, otherwise they would run out too quickly.

This is akin to GF110 and Gf100. GF100 couldn't get a fully enabled 512 core part to the public. However GF110 or the respin with some optimizations could. Not only did the respin show improved thermal characteristics it also allowed 10% higher clocks while doing it. 10% increase in clocks, 9 percent on the memory while decreasing power consumption by 15% was huge. I am not expecting this type of improvement as GF100 was crap in the first place, but if Nvidia is willing to push gk180 to a similar level of let say a 7970 ghz edition combined with the improved efficiency, it would have a product just a touch slower than a gtx 690. That is a product that is 15 percent faster(10-15% more clocks, 7 percent more cores, 17% more bandwidth) than Titan while not using anymore power. But the only way to achieve all this is they use gk180.

Cutting out the DP is bound to save some power and allow for higher clocks. Just look at gf100 quadro 6000 and gtx 480. Quadro 6000 only had clocks of 574mhz while the gtx 480 was clocked at 700. The same thing repeated it self sort as a I mentioned early with k20x and gtx titan.

I am not saying its going to be a 2880 shader part or even 2688 shader part by the way. I have no clue the specs besides it being faster. If we get a leak that shows one of the fore mentioned figures, its gk180 based. If its 2496 or in the worst case, 2304 shader based, then it gk110.

[boxleitnerb]

How would disabling DP save power? The DP-units are dedicated, they are dormant under gaming loads, aren't they?

I heard that GK110 has indeed some problems when having all 15 SMX enabled, due to the cache system. My source was very vague on this, so I don't know specifics.

780 Ti will have 2688 shaders at least, that much is certain. Only questions left are if it has even more and where the clocks are at.

Btw the 902 MHz of the K6000 is only boost, base clock is 797 MHz:

NVIDIA Quadro K6000 Graphics Card based on the GK180 GPU
Core Count: 2880
Base Clock: 797 MHz
Boost Clock: 902 MHz

The Quadro K6000 is the first NVIDIA Professional solution to incorporate Quadro Boost, a mechanism that automatically maximizes application performance while staying within the specified power envelope. For workloads that do not reach the allowed power level, the GPU clock is automatically increased to "Boost Clock" in order to leverage the remaining power budget for additional increased performance. The GPU will always try to reach the higher "Boost Clock" in order to maximize application performance.
The Quadro Boost feature will disable "Boost Clocking" in the following scenarios:
If clock jitter sensitive "Workstation features" like Sync, SDI, or SLI is enabled, the Quadro K6000 will clock to Base Clock automatically.
If the user manually selects the "Prefer Consistent Performance" Control Panel option to lock explicitly to Base Clock (can be useful for performance tuning of applications during development).

http://h18000.www1.hp.com/products/q...a/14708_na.PDF

So (K20X -> K40/K6000):
732 -> 797+
2688 -> 2880

I wouldn't expect more than 900 MHz base clock for the 780 Ti with 2880 shaders.

[PedantOne]

780 Ti has little bit higher then Titan performance, less memory, price will be higher then GTX 780 (little bit). In detail, 780 Ti IS TITAN with 3GB and higher clocks, BUT this puupy will be available in custom versions! What about (780Ti) Titan Lightning?

[boxleitnerb]

So this time around absolutely no fully unlocked version with 2880 cores for GeForce?

[PedantOne]

We can still hope for that, i smell something with BLACK cooler here

[[XC] Oj101]

And I smell some 6 GB non-reference cards, but they won't be the cheapest.

[[XC] Oj101]

I might also be smelling 6 GB reference cards, hmmm smells good

[aussie-revhead]

780 Ti has little bit higher then Titan performance, less memory, price will be higher then GTX 780 (little bit). In detail, 780 Ti IS TITAN with 3GB and higher clocks, BUT this puupy will be available in custom versions! What about (780Ti) Titan Lightning?

I'm so annoyed for buying a GTX780 Lightning at launch now, only two weeks old and its outdated already .... watch for pricing on the 780Ti Lightning, if its faster than Titan then it will probably cost more too ....

[PedantOne]

I think 780 Ti Lightning will not be here that soon, in two/three months maybe

[Lanek]

I wouldn't call it panic, more like a natural response.

Im even sure they had many solution available well before last weeks. In general AMD, Nvidia tests different solution or simulate them just for choose the sku available, so they nearly know what they could obtain from different variant, they then decide between TDP, clock speed, temp, noise, binning, availability and ofc price..

They could have launch a 780 allready with the same amount of SM of the Titan, with 3G..

[[XC] Oj101]

Yup, but it IS coming.