David Kanter (dkanter@realworldtech.com) on 3/23/12 wrote:

>Please do not call execution units cores. That's Nvidia's rather confusing (and wrong) marketing-speak.

Sure, let's stick to the author's terminology.

>>Following
>>these rumors, it seems that this future GPU will be >essentially two FK104 merged into one chip.
>
>I don't see how that is a contradiction.
>
>I was discussing the bandwidth from shared memory, which is really a reflection
>of the core design. There's no reason they cannot have fewer execution units per
>core (or perhaps more shared memory/core) and simultaneously scale up the number of cores.
>
>If you believe they are going to have a ~500mm2 design, I think it's more reasonable
>to expect about ~1.5X in overall FLOP/s.

I work neither for nVidia, nor for any of their affiliated partners, so I have no knowledge of what they will release. I can only speculate.

From what I read, nVidia chose to decrease B/FLOP in order to decrease energy consumption per execution unit and allow for higher execution unit density on the chip. I find it hard to believe that nVidia will release a Kepler chip with 2x the number of execution units of GK104, which will AT THE SAME TIME radiate more energy per execution unit and require more space per execution unit. And all of this is supposed to happen less than half a year from now (this August, allegedly).