My take on the matter is that there are good points and bad points to the two approaches (AlphaSim and Smooth Random Field = SRF).

AlphaSim good point: doesn't rely on asymptotic formula for tail probability in the limit of large smoothness. When the smoothness is more than (say) 3-4 voxel dimensions, the asymptotic formula is probably good -- for smaller smoothness, only a simulation approach like AlphaSim can give semi-accurate cluster p-values (IMHO).

AlphaSim murky point: assumes the statistics are equivalent to p-values from a jointly NORMAL random field, whereas that isn't quite true (the statistics are usually from a jointly Student-t distributed field). How much difference this makes is unclear to me. The SRF approach uses asymptotic formula derived by the late lamented Keith Worsley for each case of the underlying distribution -- this issue can be considered a good point for SRF vs. AlphaSim. (Of course, in principal AlphaSim could be modified to use different distributions, but that would be something like work.)

AlphaSim bad point: it is slow and not easily encoded into a formula -- so you can't just set a threshold and get a p-value back right away.

AlphaSim good point: it allows for the actual geometry of the brain mask in its simulations. This is probably a very minor point -- that is, it treats the edge of the brain more correctly, but how much of the brain is at the edge?

The real issue (again, IMHO) is the smoothness point. If one is smoothing the hell out of the data, then the asymptotic formulae used in the SRF method are accurate and much easier to use than AlphaSim. If one is NOT smoothing so much, then AlphaSim should be superior. At the cost of having to run the damn thing for every case, of course.