Computing the registration metric (how well the two images are aligned) at each iteration is a significant portion of the registration. The metric is evaluated at each sample point. Therefore, the mor samples you use, the longer one iteration will take.

However, if you use too few samples then your registration accuracy may suffer (not converge to global optimum, not find the optimum accurately). Also, if you use too few samples then the registration may need more iterations to reach convergence.

So, you need to find a good number samples, which ensures convergence and accuracy, but does not take too long time. Usually you determine this number for a registration problem by trial and error.

Since in BRAINSFit, metric is computed on the CPU, the higher the clock rate is, the faster the registration will be. I think metric computation in BRAINSFit is multithreaded, so CPUs with more threads can compute the registration more quickly.

You can also speed up the registration by using an multi-level scheme (first compute a coarse registration and then refine). BRAINSFit may be able to do this, but Elastix (in SlicerElastix extension) is configurable more flexibly, so you can set up sophisticated multi-resolution schemes for all kinds of registration problems.