Above is correct.
It varies by how much PhysX calculations are being done, but I've seen several tests suggesting you'd want a card at bare minimum 50% as powerful as the main card.
In one of the games tested the much weaker card gained 11% to the frame rate and having two identical GTX580's gained 15% (SLI on the other hand with 2xGTX580 in that game got over 50% improvement).
The WEAKER the dedicated GPU the less it can help and more likely it is to make things worse, and unfortunately you'd have to test EVERY GAME with PhysX on the dedicated card and on only the main card to see if it helped or hurt.
When the dedicated GPU is too slow the main GPU and rest of system just sits tapping their virtual fingers waiting for it to finish...