Isn't giving GPU computational power this way extremely inefficient? Where can I read more about implementation of this with more technical details?