See the discussion at the bottom of [Impulse response figures not show up) as well as the Appendix B7 at sites.google.com/site/bornecon/research/Born_Pfeifer_Policyrisk_Appendix.pdf?attredirects=0
Short summary: either you need really a lot of replications or you use IRFs relative to the stochastic steady state/ergodic mean in the absence of shocks as described in our Appendix (which is exactly what Risk Matters did). Also note the terminological confusion. Risk matter actually computes IRFs not as deviations from the ergodic mean, but from a different concept best termed "ergodic mean in the absence of shocks " or “stochastic steady state”. They are deterministic and you don’t need replications to make them smooth.