If you can find a 260 or 460 still somewhere, those would probably be optimal. Aside from that, a 450 or 550 is not a bad choice to pair with a 580.
From what I've seen around here, a lot of people have made up their mind's about using dedicated physX cards based on really small samples of data. Like they look the results of using a dedicated card in one game or benchmark and decide that this is 'representative' of the total population.
In reality, across the full gamut of possible scenarios, there's a pretty wide range in levels of 'usefulness' when it comes to using a dedicated card. And the most important 'deciding' factor is this: How difficult is the graphical load ASIDE from the PhysX component, relative to the power of the card(s) you're using for rendering?
If the game/bench is basically really easy to run in terms of rendering (Fluidmark provides an extreme example of what I'm referring to, but so do many of the older games that support HW PhysX), then yeah ... a dedicated card doesn't make any sense, and may even lower your performance if it's significantly slower than your main card.
OTOH, you take a game who's rendering component is actually really demanding (like, say, Batman AC, maxed out at 1080p in DX11 with full AA), then you're in a scenario where you have no 'extra' GPU power to spare to do the physX calculations.
In such a case (assuming you've enable PhysX in the game) practically ANY physX card is going to provide you with a performance boost over not having one at all, and you're likely to see some pretty nice perf scaling with the quality of the physX card that you deploy (up to certain point, at least).