Internally it is implemented by spliting the number into an array of conveniently sized chunks (traditionally called limbs). That might mean either the machine

Internally it is implemented by spliting the number into an array of conveniently sized chunks (traditionally called limbs). That might mean either the machine word size for optimized assembly implementation on many common platforms or one bit less than native word size for portable C implementation (and architectures without carry flag) or some fixed number like 15 bits (IIRC mbedTLS has or at least used to have hardcoded limb size of 15bits in its bignum implementation).