This work provides an in-depth computational performance study of the parallel finite-difference time-domain (FDTD) method. The parallelization is done at various levels including: shared- (OpenMP) and distributed- (MPI) memory paradigms and vectorization on three different architectures: Intel's Knights Landing, Skylake and ARM's Cavium ThunderX2. This study contributes to prove, in a systematic manner, the well-established claim within the Computational Electromagnetic community, that the main factor limiting FDTD performance, in realistic problems, is the memory bandwidth. Consequently a memory bandwidth threshold can be assessed depending on the problem size in order to attain optimal performance. Finally, the results of this study have been used to optimize the workload balancing of simulation of a bioelectromagnetic problem consisting in the exposure of a human model to a reverberation chamber-like environment.