The algo is 2mb scratchpad.

That means it needs 2mb of CPU cache per core/thread.
You probably got 16mb cache, so 16 / 2 = 8

You might get a tiny boost using more than 8 cores, but not that much.

EDIT: correction, TUBE is 4 MB scratchpad...so you might see most efficient hash/s at 4 cores with 16mb cache, more cores will add a bit, but not that much
Its all about scratchpad and cache simply