Unlikely to happen. Reference counting has poor interaction with the CPU memory hierarchy: It makes all objects larger, making the working set correspondingly larger, meaning your data cache is less effective. It makes code larger because of the extra ref counting twiddling code. Larger code means a less effective instruction cache, as well as being slower because of all the extra operations performed. Reference counts must be frequently modified, and they are scattered all over memory (next to each object), so you get poorer locality and more atomic writes which has all sorts of negative implications for cache-coherent SMP architectures.
So while you may not notice the penalty of reference counting since it is spread out over every operation, it's unlikely that it is better than proper GC for most applications. Maybe just for ones which are extremely sensitive to latency, yet don't mind running much slower overall (hard real time? aeronautics and space?).
Gc pauses suck the "magic" out of interactive programs. The great thing about obj-c is that it I trivial to have a c array of vanilla c structs, so your performance sensitive code is still slim and hot. (size and cache locality.)
My iOS audio programming mixes c for audio processing and obj-c for ui and control quite effectively.
I didn't say that GC was free ... malloc has overhead too.
For interactive programs, I would recommend two things to avoid (or hide) the pauses. Firstly schedule calls which run the GC when it won't be noticed: between frames in an arcade-style game, or just before accepting user input in a turn-based game. Secondly, size the minor heap large enough so that all computation between the scheduled GC runs can happen without invoking a minor sweep (but not too large that the minor heap is wasting memory -- some experimentation and tuning required).
I've written a couple of interactive games in OCaml that used this strategy and avoided visible GC pauses.
Last I looked, it was a bit hard to follow, but iOS seemed to be storing reference counts separately from the objects. I think you could do (and maybe iOS does?) this on a per-thread basis, to reduce the need for locks.
So while you may not notice the penalty of reference counting since it is spread out over every operation, it's unlikely that it is better than proper GC for most applications. Maybe just for ones which are extremely sensitive to latency, yet don't mind running much slower overall (hard real time? aeronautics and space?).
As usual you'd need to measure it.