What is needed to make a GPU fully fed by CPU?

Comparing their flops/s is enough?

For example, if i could manage to make a very old(pentium-3) CPU with one of Nvidia-Fermi GPU, it would not be able to fed the gpu with data per sec. What is the criteria to fit CPU to GPU exactly when OpenCL or some similar work needed?

Of course RAM and BUS will be chosen in a similar way but how exactly?

Assume each GPU-core will calculate a sqrt and a division and an adding for 100 times for every itearation.

Thanks.