However, when Nvidia Tesla-100 is running a program with the arithmetic intensity 0.25 or lower (e.g., 2 operation per each DP number loaded from the memory, like SpMV product or, more generaly, HPCG), then the GPU performance could not be higher than 225 GFlop/s.
In fact, in the Summit supercomputer that uses Nvidia Teslas, the actual performance when running the HPCG is about 106 GFlop/s per one GPU.