Its mainly used for the DP floating point that isn't limited artificially. The titan supercomputer at ORNL uses an older version of these. So heavy compute.