Yeah, atomic RC hurts a lot (5x-10x slower, based on the measurements I've seen -- Hans Boehm of Boehm GC fame did a presentation on this that you can find if you're interested). (This is a problem with Objective-C reference counting -- they use atomic ops for all reference counting I believe.) There's some work on mitigating its cost with per-thread reference counts arranged in a tree or whatever, but it's still tough.

The alternatives are either (a) avoiding shared state; (b) strong lifetime management such that you make sure one parent thread owns all of the data that its child threads are going to use; (c) concurrent GC; (d) pretending it's not a problem like Go does and doing stop-the-world GC.