We developed the energy and its gradient for the self-consistent-charge density-functional tight-binding (DFTB) method, combined with the fragment molecular orbital (FMO) approach, FMO-DFTB, including an optional a posteriori treatment for dispersion interaction, and evaluated its accuracy as well as computational efficiency for a set of representative systems: polypeptides, a DNA segment, and a small protein. The error in the total energy of FMO-DFTB versus full SCC-DFTB was below 1 kcal/mol for the polyalanine system consisting of about 2000 atoms partitioned into fragments containing 2 residues, and the optimized structures had root-mean-square deviations below 0.1 Å. The scaling of FMO-DFTB with the system size N is only marginally larger than linear [O(N(1.2)) in the worst case]. A parallelization efficiency of 94% was achieved using 128 CPU cores, and we demonstrate the applicability of FMO-DFTB for systems containing more than one million atoms by performing a geometry optimization of a fullerite cluster.