[citation][nom]stm1185[/nom]So those leaks showing Kepler having 1536 Cuda cores could actually be true then![/citation]
GK104 is supposed to have 1536 cores WITHOUT hot-clocking, making them effectively a LOT slower than a Fermi CUDA core. Besides, GK100/110 is supposed to have more than that.
[citation][nom]KillerBees[/nom]You idiots need to read further and understand that this card is only 10% faster in Battlefield 3 vs HD7970. That is a FAIL!!! Here is proof for you Nvidia fanboys that this card is not that great: http://wccftech.com/nvidia-kepler- [...] er-hd7970/[/citation]
That card clearly has a GK-104, NOT a GTX 680 GPU. 680 will likely have a GK-100 or 110. I'm not too good with Nvidia GPU names so I'm not sure which it will be, but it's probably one of those two. That is something like a GTX 660 or 660 TI. GK-104 is supposed to have 1536 cores, NOT the high end GPUs in the 680, 670, maybe 660 TI. Of course, Nvidia might change names around too, so who knows?
Also, GK-104 is about the same size as Tahiti, so it is fairly similar in performance. AMD and Nvidia have similar performance per square mm of die. Considering that, and the increased power usage over the 7970, the 7970 and this GTX whatever probably overclock to about the same performance. The Nvidia card listed at that site is about 10% faster and uses about 20% more power, seems fairly reasonable if the main difference is clock rates (remember, clock rates increase power usage exponentially while increasing performance linearly).
Basically, the 7970 and this Nvidia card can probably run at similar performance while using similar amounts of power. However, it clearly states that the card is only UP TO 10% faster and is usally slower in 3Dmark, so it probably doesn't beat the 7970 anyway, in which case Nvidia could have failed. Then again, all of this is assuming that this link of yours isn't fake anyway. Since it really has little to say that even makes me assume it's true (it partially contradicts what Nvidia has been saying), I'm leaning on it being BS.
For example, it says that this card has hot-clocking despite Nvidia claiming that Kepler doesn't and the power usage of 1536 CUDA cores that still have hot-clocking would be through the roof EVEN on the 28nm node. Then we see it has much less memory bandwidth than the 7970. The picture also has no defining features to tell us anything about the card save for it having a fan, being black, and it has Nvidia on it. For all we know, it's some old card. Nothing on the card implicates otherwise.