These generators are based on an AlgolW program distributed by Donald Knuth to his CS144b class in 1975. They use the additive random number generation algorithm that is recommended in the second edition of Seminumerical Algorithms by Knuth, and that was the subject of a Ph.D. thesis by John Reiser of Stanford ("The analysis of additive random number generators", STAN-CS-77-601, March 1977.) The additive generator has several advantages over a standard linear congruential generator (lcg):
The sequences that it generates are more "random" in several ways. With an lcg generating a sequence a(n), the related sequence a(n) mod 2 has period 2, a(n) mod 4 has period 4, and so on; thus care must be taken not to derive any results primarily from the least-significant bits of a(n). An additive generator produces numbers whose bits are more uniformly random. An lcg for a short wordlength has a short period (for example, a sequence of 16 bit cardinals must cycle after 64k terms); an additive generator may have a very long period even for short wordlengths. Many lcgs (including IBM's notorious RANDU) generate sequences that are very poorly distributed in two and three dimensions (you might see this by plotting consecutive pairs or triples of terms in euclidean space and noting the patterns); additive generators do not have this problem.
The generator relatively easy to transport from machine to machine, and can be made to generate the same sequences on each machine. This can be difficult with lcgs because the sequences they generate usually depend on the way a machine handles overflows (in particular, on the wordlength.) An additive generator can produce a random sequence without ever causing an overflow.
The generator is relatively fast because it does not use multiplication, which is a time-consuming operation on our present machines.