Is it possible to assign only n cores to perform the kernel computation to avoid freezes my ubuntu desktop? my only option is access via ssh an reboot the pc
PD: I'm using pyopencl 2012.1 with a Radeon HD 6450 and the drivers included in ubuntu 12.04 (fglrx_8.960).
Your best bet is to get a low-end card and use it for display.
Use the high end for computation.
There is no way yet to partition the GPU.
Device Fission is an OpenCL capability -- but at the moment AMD supports only the CPU device for it.