%CPU shows the cputime/realtime percentage in the period of time between updates. For the first update, a short delay is used, and top itself dominates the CPU usage. After that, top will drop back, and a more reliable estimate of CPU usage is available.
The SIZE and RSS fields don't count the page tables and the task_struct of a process; this is at least 12K of memory that is always resident. SIZE is the virtual size of the process (code+data+stack).
Keep in mind that a process must die for its time to be recorded on its parent by cumulative mode. Perhaps more useful behavior would be to follow each process upwards, adding time, but that would be more expensive, possibly prohibitively so. In any case, that would make top 's behavior incompatible with ps .