The GGNFS siever has a check that its large primes are no more than 33 bits, but as far as I can tell there's no underlying reason for this; the large primes come out at the postfactoring stage and at that point you've evaluated f(x,y) as an mpz_t.
The yields from using 33bit large primes to sieve a C91 are somewhat impressive; about 600,000 relations per kiloQ and 3k relations per second. Of course, the winnowing from singleton removal is equally impressive; sieving overnight and will try to see if I get a matrix in the morning :)
(for example, it is not until 4.9 million unique input relations and 9.7 million unique input ideals that I first get an ideal surviving singleton reduction)
Last fiddled with by fivemack on 20090921 at 22:44
