My test code is here: http://chezphil.org/tmp/boost_mp_review.tbz2
Just a couple of updates: 1) Your test code doesn't actually test fixed-width integer types anywhere as you have: #ifdef USE_MP_INTN_NOET typedef int64_t i64; typedef boost::multiprecision::mp_number< boost::multiprecision::cpp_int_backend<>, false> i128; #endif Should be: #ifdef USE_MP_INTN_NOET typedef int64_t i64; typedef boost::multiprecision::mp_number< boost::multiprecision::cpp_int_backend<128, true, void>, false> i128; #endif Of course the types are pretty similar, and it only saves a few percent - relative times on Win32 were: Your custom int128 1.9. mp_int128 2.7. cpp_int (no ET's) 3.2. 2) Some more comparisons with gmp shows expression templates help much more in that case: proposed mpz_int 100 proposed mpz_int (no ET's) 143 GMP's mpz_class [1] 140 Cheers, John. 1) With apologies to Marc Glisse I used the last released mpz_class, still haven't looked at his improvements in GMP development :-(