Energy drift - usually damping - is substantial for numerical integration schemes that are not symplectic, such as the Runge-Kutta family.
[1][2] The accuracy of the energy conservation for the true Hamiltonian is dependent on the time step.
Energy drift is similar to parametric resonance in that a finite, discrete timestepping scheme will result in nonphysical, limited sampling of motions with frequencies close to the frequency of velocity updates.
For example, cutoff schemes for evaluating the electrostatic forces introduce systematic errors in the energy with each time step as particles move back and forth across the cutoff radius if sufficient smoothing is not used.
Particle mesh Ewald summation is one solution for this effect, but introduces artifacts of its own.
Errors in the system being simulated can also induce energy drifts characterized as "explosive" that are not artifacts, but are reflective of the instability of the initial conditions; this may occur when the system has not been subjected to sufficient structural minimization before beginning production dynamics.
However, it has been shown that long microcanonical ensemble simulations can be performed with insignificant energy drift, including those of flexible molecules which incorporate constraints and Ewald summations.
[1][2] Energy drift is often used as a measure of the quality of the simulation, and has been proposed as one quality metric to be routinely reported in a mass repository of molecular dynamics trajectory data analogous to the Protein Data Bank.