Depending on what you are trying to simulate, you may have way better luck with a phase domain model.  Convergence problems usually occur with voltage domain models if you are using an RF simulator. The VCO reveals the difference between the two models. In a voltage domain model, the VCO outputs an oscillatory waveform. In phase domain model, the VCO outputs a ramp. Phase domain models have their share of convergence problems too but they are much easier to solve than the voltage domain convergence problems. One other nice thing about a phase domain model, besides the dramatic increase in speed, is that you can move the VCO and reference integrators into the PFD, thereby forcing the DC analysis to pre-charge the loop filter to the correct operating point.  I've implied here that you simulate the phase domain model with Spectre or Spice, not SpectreRF. Since there is no hidden state problem with Spectre/Spice, if you doing a fractional N, you can use the real sigma delta modulator, which is guaranteed to have hidden state.