The more interesting fact is the card will have 512-bit bus, resulting in 384 GB/s memory bandwidth, because 6GHz memory speed.
Impressive :)

WhyCry you mixed up the lines for ROPs and TMUs.

In Kepler (GK104), the ROPs are broken up as 4 ROPs per 1 SMX cluster. Since GK104 has 8 SMX clusters, you get 32 ROPs. Assuming a similar scenario for GK110, we’d get (2688 / 192) = 14 SMX clusters or 56 ROPs. The full GK110 has 15 SMX clusters and 6 memory controllers, with 64-bit per controller, for a total bandwidth of 384-bit. How can you have 512-bit when GK110 has a maximum of 384?

http://www.anandtech.com/show/6446/nvidia-launches-tesla-k20-k20x-gk110-arrives-at-last

Looks more like they directly copied 690 clock/memory interface specs.

The full GK110 has 6 memory controllers per this diagram:
http://images.anandtech.com/doci/5840/GK110Block.png

How did they get 512-bit bus? I think that retailer incorrectly copied the specs for GTX690 for the Titan or it’s a placeholder. Laughable Source beat me to it.

Maybe Nvidia won the lottery and they do not give a damn about expensive GPU die space, so they decided to widen the front and sides memory bus for the heck of it!

I remember rumors of 768 CUDA core 580, 2304 CUDA core 680, etc. Always overhyped!

You are right. I didn’t spend much of the post about this, because I highly doubt that’s even possible. However, I wanted to note that there is such a ‘strange’ interface listed, so people would not be amazed how I missed that :)

Why $1599? It supposed to be $899? Why they sell it with $700 (!) higher price? I wonder what is the prime cost?
Such pricing policy can lead nVidia to the loss of the market share. They’re should spend less money for the promotion or even reduce staff to be able to reduce prices.

The funny thing is if performance of this card in true minus the bus width what would be left for the GTX 780?? I mean a refresh cycle has never ever produce this kind of jump in performance before, So I assume that the 700 series will be based somewhat on the GK110 or some variants. It has too or else Titan will be the fastest card ever until the arrival of Maxwell!

$1,599 sounds like a placeholder price. A lot of theses stores abroad sell at inflated prices compared to the US. Either way, the Titan’s purpose is not to sell millions of units. It’s a halo card to drum up interest for the NV brand by reclaiming the performance crown. It’s going to serve a similar purpose to the GTX690. Those are low volume seller cards to begin with.

I have a new alternative theory. Since NV is selling K20X parts for more than $3,000, wouldn’t it make sense for them to resell parts that couldn’t make Tesla instead of throwing them out? Then the chips that didn’t yield full 14 SMX clusters could be taken and overclocked. You’d end up with 13 SMX 2496 Titan but clocks at 1019mhz. This way you still have a very fast card and the TDP goes up to 250W. Since 732mhz + 2688 CUDA cores gives us a TDP of 235W, could 13 SMX @ 1019mhz at 250W TDP be a better trade-off for Nvidia to take? The projection for 1019mhz and 2688 CUDA cores at only a 15W TDP hit seems unrealistic, no?

It’s a good theory, however too many source have claimed it has 14 SMXs. I would rather stick to the story with 384-bit bus and lower sub-GHz clock. As for the power consumption I bet this is the very last of their concerns with this card. Whoever is going to buy one, doesn’t care about TDP.

Guys and what is the 6GB memory, where they can do or stick)

6GB of VRAM is to flex your e-Peen after spending 9 bills :)