On
http://www.hogranch.com/mayer/gimps_timings.html i read about MLucas 2.7b
Quote:
Supports the same nonpowerof2 runlengths as Prime95, i.e. (1,3,5,7)*2n, plus a full set of 4 additional intermediate radices (9,11,13,15)*2n. This allows every interval between adjacent powerof2 runlengths to be broken into eight roughly equal pieces, and runtimes to increase very gradually as exponents get larger.

Are these planned for a future version of prime95? From the timings it looks like a 10M digit test can be about 10% faster with a 1680K FFT