I think it's reasonable to quote single precision (32bit) FLOPS, given the origin of the term - but quoting half-precision figures is taking the biscuit. Presumably NVidia's benchmarketeers will be quoting quarter-precision FLOPS next time round to fool people into thinking their next gen is twice as quick as the P100.
If pressed I would speculate that some real-time signal processing app out there can make use of 21 Thalflops, I'd be interested to hear what kind of apps folks think those halfFLOPs will be good for. :)