I checked a rewrite of the cell-host grid system into SVN a few weeks back. I’m going to make a branch of the current baseline and re-integrate the new grid system code into that, instrument it up with unit tests, and see if I can’t have it fully working by tonight, then run it through it’s paces tomorrow.
Part of the new grid system is slightly better optimized vis-list building on the host. More importantly, the new mechanism is encapsulated in an entirely thread-safe way, so I can potentially parallelize it, which means I can be a little more methodical inside the processing routines.
The cell hosts themselves are dual core-2 xeons (4 physical cpu cores), but we get very little work out of more than 1 core at a time :(