Good catch, I think the 30x came from including the advantage given by tensor cores at all and not just lower precision data types.
Good catch, I think the 30x came from including the advantage given by tensor cores at all and not just lower precision data types.