Since a HCC1-GPU task reserves 1 CPU core thread by default for the whole duration of the task, on e.g. a quad, 3 CPU jobs can run at most, and one HCC-GPU task has distinct CPU (c) and GPU (g) processing segments in the order of c-g-c-c-g-c, there will be idle sections in the order of g-c-g-g-c-g. The spare time can be filled by configuring app_config to allow running of multiple HCC-GPU tasks off from one CPU core. The filling of the idle gaps is coincidental, i.e. running offset for maximum utilization. Since each task is discreetly running in it's own memory segment, there's no way to manage this [ground up redesign of BOINC requiring to get that], trial and error is needed to find the best number to run concurrent without sliding into a negative productivity state... too many causes so much overhead, you'd do less work rather than more.

Follow prior threads in GPU on HCC1 forums for discussion of *how to*. Several Help topics have been drafted, but those likely wont be released until WCG formally adopts the 7.0.xx client.

P.S. To repeat, the WCG servers remain to be oblivious to the existence of the app_config.xml. A *future* client will be enabled to transmit the settings back to the server so it can anticipate for instance work assignments such as how many to allow in progress for a host for an individual science. 7.0.42 is *not* doing that yet at times of a scheduled or manually forced update.