The comparison is completely fair, because on RISC-V there is no better way to generate the carries required for computations with large integers. You cannot generate a carry with a 64-bit addition, because it is lost and you cannot store it.

You should take into account that the libgmp authors have a huge amount of experience in implementing operations with large integers on a very large number of CPU architectures, i.e. on all architectures supported by gcc, and for most of those architectures libgmp has been the fastest during many years, or it still is the fastest.