Moore's Law in action- 13TeraFlops on a single card
http://techfrag.com/2016/06/10/nvidia-gtx-1080-ti-titan-rumors-gp102-gpu-delivers-13-tflops/
The GP102 GPU will probably include 3840 CUDA cores, but what’s more interesting is that it is rumored to deliver performance equivalent to the GTX 1080 2xSLI, which means the next Titan X should be 50% faster than a single GTX 1080. Believing this, the GP102 GPU must be capable of delivering 13 TFLOPS of compute performance provided it is clocked at 1733 MHz, the same boost clock as the GTX 1080.