Nah , that's impossible.
The only explanation is IMO that something is wrong/suboptimal when running only one thread.Which is weird since one thread should have all the L3 for itself.
Moreover , one thread has both IMCs working for it.
In the multithreaded scenario , data BW from both L3 and memory is shared => welcome to Amdahl's law.
http://en.wikipedia.org/wiki/Amdahl's_law
Bookmarks