Also, LLVM static analyze is so good at figuring out your code path that it could effectively retain and release your memory for you.
Cyclic reference is its only flaw but its something that can easily solved through programmer awareness.
ARC may be less likely to get memory paused on release, but ARC doesn't compact free memory, and therefore fragmentation can cause object allocations (malloc) to become more expensive, possibly leading to pauses.
If you use non-copying GC that doesn't compact (ARC doesn't compact), then there are low pause GC algorithms out there that give pretty good bounds (e.g. 5ms pause).
I'd like to see actual benchmarks of both ARC and say, mark-and-sweep on a mobile device rather than speculation and opinion, and both benchmarks must get the same static compilation treatment. That is, if escape analysis tells you that the lifetime of the object is bounded by the current stack frame, then allocate it on the stack, and don't penalize the non-reference-counted GC by allocating 100% of everything on the heap.
A fair few years ago I worked in the murky world of J2ME games, and a common pattern was byte[] _variables = new byte[1024] so you could ensure there were no GC pauses.
That's not the same as ARC, though. The Java equivalent to ARC would be something like having an array of volatile reference counts to go with your variables and fiddling with those at most accesses.
It's not like tons of games haven't been shipped with non ref-count GC. Minecraft being the most famous, but games on Unity3D/Mono can have non-ref count GC. Lua is shipped in tons of game engines and uses classic GC.
Regardless of whether you are using automatic GC, or if you are using malloc/free, you can't write a high performance game without carefully working around memory issues. Even C/C++ games like Max Payne have shipped with frame hiccups caused by poor malloc/free behavior that had to be fixed with custom allocators.
If you're writing a game, you have to pay attention to what you're doing. But do non-framrate limited games and regular apps need ref-count GC? I suggest no, they do not.
Practically the first piece of advice usually given when someone asks "how do I make my Android app non-laggy" is to avoid allocation wherever possible; even a single frame drop due to a GC pause when the user is scrolling a list, say, can be noticeable.
Low pause GC is possible. See http://openjdk.java.net/jeps/189 or Zing/Azul (absolutely pauseless GC or so they claim)
What's FirefoxOS doing for it's GC?