The prominent V100 feature it’s tensor cores and DNN applications. For applications that were already efficient with Pascal architectures, I suppose that the price increase is more difficult to accept.
Back to the initial question, I forgot to mention the approximate hard coded maths functions (exp sin sqrt…) that can lead to spectacular speed ups compared to IEEE soft implementations.
Last, I speculate that the new RTX cards could have new spectacular scientific applications in Monte Carlo Transport, or places where ray tracing is important.