26 Jan 12 // Latest info from a group of websites including xbit, fudzilla etc..
Kepler GK 104, looks like it will have a 256bit bus, with 2G Ram, a 225W TDP, and an expected $299 price tag. Looks like it is going to be the GTX560/70 successor.
Other sites estimate that there will be 768 shaders to achieve the expected 2 Teraflops, but it is unclear to me if that speculation is the GK-104 (midrange), or the full thermal envelope of the final top Kepler chip.
**Some estimates are projecting as many as 768 CUDA cores or "well above 2 teraflops" of raw performance
Introduced technology advantages would include 64-bit floating operations of 4 times Fermi's capacity per watt. Another would be the virtual memory space shared between CPU and GPU.
Additionally, noticed this wikipedia entry, not sure what to think of it, or where that info came from other than the supplied Toms Hardware link:
http://en.wikipedia.org/wiki/GeForce_600_Series
http://news.softpedia.com/news/Nvidia-GK104-Kepler-GPU-May-Be-Priced-at-299-230-248594.shtml
http://semiaccurate.com/2012/01/23/exclusive-and-the-nvidia-keplergk104-price-is/
**
http://www.fudzilla.com/home/item/25636-nvidia-28nm-gk104-gpu-specs-revealed
Additionally,
here is some decent math figured out based on TDP and die shrink area, to figure out a possible ratio for clock vs number of shaders:
Look at the reference models of GTX570 and the 448-core GTX560 Ti, all frequencies and Vram sizes are a match, except one has 32 less CUDA cores and accounted for from consuming 9W less power typical load. Doing the math reveals the die of the 448-core GF110 consumes 126W of power while the 480-core GF110 consumes 135W power, and the GK114 384-core ends up down at 115W taking into account the higher frequency. A shrink to 28nm takes 51% of the area out of a 40nm die, which is approximately porportional to wattage loads, so a 28nm 384-core would end up 62W draw. This would make a 28nm version of GTX560 Ti at 110W -- same as GTX550 Ti, maybe it could get called "GTX650 Ti".
But if we double the cores to 768 then we're right back up to 170W card, so what accounts for 225W? The GPU would have to have either 1024-cores at 1770Mhz, or some other combination with higher frequencies or less cores. Or the Vram frequency would increase, but the Vram isn't the majority of the other circuitry, so it may not account for that much.
My own discussion about this:
Of course this is rough math, but some things make sense. The change from 40 to 28nm does translate to a 51% decrease in area, calculated here ( 28 x 28 ) / ( 40 x 40 ) = 0.49
What the person in the quote above didn't account for, is a chip shrink. If the chip were to be the exact same size as Fermi, then yes, a 51% increase in total shaders would happen. Of course, this is a new chip, so the size might change. So this math is not irrelevant, but too many factors are left out (frequency, # of shaders, and die size) to get a final conclusion. If any of those values would come out, then we could indeed play around with some more concrete figures.
Another thing could be the fact that if the final chip is 1024 shaders, the 660 would be 768 shaders, and the chip size would be very similar to fermi's, then you could conclude that the 170watts simply means that 25% of the card wasn't used aka its the same chip as the GK-110 but with a quarter disabled. ~170 x 1.25 = ~212.5... this brings it close to 225, but then you can also assume that GK-104 could be clocked slightly higher to compensate for the lower shader count (commonly used in the midrange sector), so 225W would make sense at this point
for this guy ^
