ndbmtd runs slower than ndb (no replies)

Below are my 3 readings of a LOAD DATA query of 2 M rows.

Ndb Ndb:Diskless Ndb:Diskless + MT (12 threads)
547 s 503 s 758 s

With 12 threads, I was expecting some improvements in the LOAD DATA. But I got slower performance instead. How to explain this?