The architecture of the NVIDIA GTX 780 Ti core, the GK110 is an evolution of the GTX680's GK104. Both cores are based on Kepler architecture and both are built upon the 28nm ASIC. However the GK110 features 7.1 billion transistors compared to the 3.54 billion found in the GTX680 and the 4.4 billion in the AMD HD7970. The CUDA core count has also been increased to a 'full fat' 2880 from the 2688 used on the GTX TITAN and from 1536 of the GTX680.
The GK110 has an additional GPC to the four found on the GK104 and also and additional SMX in each GPC making 15 compared to the GTX 680's 8. There are 6 64-bit memory controllers (384bit) controlling the huge 6GB of GDDR5 which is yet another massive hike over the 256bit controller found on the GK104.
Above we can see a breakdown of an individual SMX.
NVIDIA's CUDA core count on the GTX 780 Ti weighs in at an astonishing 2880. This is a huge undertaking as every SMX needs to be perfect. Previously at least a single SMX would be disabled and thus the silicon would then be 'binned' as either a GTX TITAN (14 SMX - 2688 CUDA cores) or GTX780 (12 SMX - 2304 CUDA cores). The GTX 780Ti uses all 15 SMX so understandably, this will be an expensive card to produce.
To combat this, NVIDIA have made the wise decision to reduce the amount of on-board GDDR5 to 3GB. The GTX TITAN has 6GB of on board memory and while this amount can be utilised under certain scenarios, 2-3GB is more than enough for the vast majority of uses. That said, while we can see the wisdom in reducing the memory size, we would still have preferred to see a 290X equalling 4GB AND a higher memory bus width which remains 384bit as opposed to the 290X 512bit. The raw grunt of the extra cores should however ensure that the GTX 780 Ti is a very capable graphics card not only for today's games but also for future titles too.