HPC II DDR controller performance

This sounds similar to something I saw a year or two ago. It only happened with the HP2 controller and not the HP1 version (altmemphy I think). If I remember correctly the bug was that the command queue wasn't deep enough to keep up with the fill rate from my masters. Try switching to the HP1 controller to see if that solves the throughput problem. If that does the trick then you would have to upgrade your tools in order to use the HP2 controller. I can't remember when this was fixed but my guess is 10.0 or 10.1 resolved it (it's fixed I just can't remember when it happened).