My understanding/finding during my experimentation on SFMT and TinyMT on Erlang from 2010 to 2012 was that the overhead of function callings themselves became the fundamental overhead of generating random numbers as a sequence with state manipulation. See my TinyMT paper Section 4.3 for the details. Quote:
While TinyMT is inherently simpler than SFMT, that does not necessarily mean TinyMT is faster for Erlang. We think this is due to the execution overhead time of each function in BEAM.
If you really need a blazingly fast random number generator, you need to consider a batch generation of 100s or 1000s of random number sequence in NIFs as a list and simply get the head element out of the list pool.