A lookup table in memory can only be accessed an element at a time, so it gets bottlenecked in the execution units on the load and address generation traffic. This algorithm uses 8-bit integers, so the vectorized version is processing between 16-64 elements per operation depending on the vector width. It's even worse if the 8-bit divide is integrated with other vector operations as then the lookup table method also has insert/extract overhead to exchange the values between the vector and integer units.

A hybrid approach using small in-register lookup tables in the vector unit (pshufb/tbl) can be lucrative, but are very limited in table size.