Despite DNA being a very important target for several proteins and drugs, molecular dynamics simulations with nucleic acids still encompass many challenges, such as the reliability of the chosen force field. In this paper, we carried out molecular dynamics simulations of the Dickerson-Drew dodecamer comparing GROMOS 53A6 and AMBER 03 force fields. While the AMBER force field presents specific topologies for the 5' and 3' terminal nucleotides, the GROMOS force field considers all nucleotides in the same way. To investigate the effects of the terminal nucleotide definitions, both force fields were modified to be applied in the two possible ways: with or without specific terminal nucleotide topologies. The analysis of global stability (rmsd, number of base pairs and hydrogen bonds) showed that both systems simulated with AMBER were stable, while the system simulated with the original GROMOS topologies was very unstable after 5 ns. When specific terminal topologies were included for GROMOS force field, DNA denaturation was delayed until 15 ns, but not avoided. The alpha/gamma transitions also displayed a strong dependence on the force field, but not on the terminal nucleotide definitions: AMBER simulations mainly explored configurations corresponding to the global minimum, while GROMOS simulations exhibited, very early in the simulations, an extensive sampling of local minima that may facilitate transitions to A-DNA isoform. The epsilon/zeta sampling was dependent both on the force field and on the terminal nucleotide definitions: while the AMBER simulations displayed well-defined B-I -> B-II transitions, the GROMOS force field clearly favored the B-I conformation. Also, the system simulated with the original GROMOS topologies displayed uncoupled epsilon/zeta transitions, leading to noncanonical conformations, but this was reverted when the new terminal nucleotide topologies were applied. Finally, the GROMOS force field leads to strong geometrical deformations on the DNA (overestimated groove widths and roll and strongly underestimated twist and slide), which restrict the use of GROMOS force field in long time scale DNA simulations unless a further reparametrization is made.