Stacked RAM is the next big thing. It will enable much faster communication between the chip and the memory and with the current 384-bit design such we saw on Kepler, the stacked RAM version can end up with 800GB/s to 1000GB/s bandwidth.
This is some significantly faster than the 288 GB/s that we saw with GK110-based products such as the Geforce GTX 780 Ti or Geforce Titan. The K20X, the fastest professional solution available today, is stuck at 250GB/s with the hope that its Maxwell successor will get slightly higher bandwidth.
The new point to point communication chip called NVLink will also help solve the memory bandwidth issue. Most algorithms in the word that are used for computation or hard tasks in graphics such as ray tracing will always end up asking for faster memory access. This is why Pascal will play a major role in the future of compute units such as Nvidia's Tesla products, or faster 4K or gaming of the future.
We saw some presentations at the GTC with a projection of Volta (now Pascal) where a scientist expects that with 1024GB per second bandwidth, Pascal could be aiming at 4000GFLOPS a second in dual precision and 12000GFLOPS a second in single precision computation.
To put things in perspective the GK 110-based Tesla K20x with 14 SMXs and 250GB/s bandwidth offers 1310GFLOPS of dual precision and 3950GFLOPS in single precision. Tesla K20x is an actual product, shipping as we speak, while the Volta (Pascal stacked RAM GPU that hits the market in 2016 ed.) numbers are projections. These final numbers will depend on the final clock and configuration of the high-end Pascal chip.
Based on these projections you can get close to three times faster computation power, and this is a huge leap that will happen some two years from now.
The future looks interesting and don’t worry, if you want to do realistic rendering and ray traced lightning effects or global illumination, you will need at least a few of these 12000GFLOPS per second cards to enjoy a real time experience. Real time ray tracing in computer games will still be a dream even beyond significant computational and memory improvements in 2016. However, this does not mean that we will not see some hybrid ray tracing solutions over the next few years.