For research institutions, Nvidia's latest Volta architecture has been available for several months as Tesla V100 for around 11,000 euros. This card contains a fully equipped GV100 chip with 16 GB HBM memory and 5376 CUDA cores. It is rumored that it costs Nvidia itself about 1,000 euros to build such a graphics card. This may also explain the price for the now presented flagship of the Titan series: The new Titan no longer costs around 1,000 dollars, but three times as much.
For the 2999 Dollar or 3100 Euro you can also get "almost a Tesla V100": The Titan V has 5120 CUDA cores and "only" 12 GB HBM2 memory. The biggest limitation this time is memory bandwidth. This is "only" 653 GB/s, while the Tesla V100 reaches a full 900 GB/s in this case. In our experience, it is the memory bandwidth that limits most complex real-time effects in video editing programs. Since a Titan XP for 1,200 US-Dollar already achieves 548 GB/s and a more affordable GTX1080Ti already comes to 485 GB/s (and is usually overclockable) the new Titan V doesn't seem to have any big performance surprises in store for video editing. With a nominal power consumption of 250W, however, it works quite efficiently for the expected performance.
Anyone who is also interested in artificial intelligence will be pleased about the special 640 Tensor Cores. These accelerate certain neuronal computational processes in hardware using cuDNN and can achieve an astonishing 110 TeraFLOPS in this field. In this domain the Titan V for 3,000 US-Dollar is a real bargain, but unfortunately no video application supports these special tensor-core optimizations so far.