The CXCL12 WUs contain relatively few small molecules with fewer atoms (about 30k) than typical long-runs (70k - 80k). The less atoms a WU contains, the faster each time step is computed (if everything else, including the scientific model is comparable). For my GTX970 it's about 2 ms per time step, compared to e.g. 5 ms for a SDOERR_BARNA5. The overall runtime of these WUs is comparable to other long runs, which means they must contain more time steps. Some CPU intervention is required after each time step, I think, or at least proportionally to the number of time steps. That's a good reason for the CPU usage of these smaller WUs to be higher :)

MrS
____________
Scanning for our furry friends since Jan 2002