GPUs are becoming increasingly important in lattice calculations, for
they combine a outstanding computational power in an economic and
compact package. In order to allow the lattice community to profit
from GPUs computing capabilities, a special library called "QUDA" has
been developed, comprising a number of kernels that accelerate by GPUs
the heaviest LQCD calculations, but surprisingly, tools to evaluate
disconnected diagrams were missing from the library. In this work we
expose our contribution to the QUDA library, and show how GPUs can
remove in an effective way, if not all, most errors derived from the
fact that we are estimating stochastically the inverse of the
fermionic matrix.