Lacking any other information, I would look into this explanation first - I have seen many experiments on cores with SMT contexts where the performance plateaus at the total number of cores, not SMT contexts.