The most optimized version I'm aware of is the RCImmix collector for Java: http://users.cecs.anu.edu.au/~steveb/downloads/pdf/rcix-oops.... It achieved performance parity with the regular Immix collector, which was the fastest (in the sense of throughput) available for the Jikes research JVM at the time. A backup collector is required for two reasons: cycle collection, and resetting stuck counts (they gain significant performance wins from keeping only a few bits for the actual refcount, for a variety of reasons: usually only a small percentage of objects need more, and they tend to be the most heavily accessed objects, so they save a lot of time by avoiding updates to them and also gain from being able to reuse objects' existing metadata fields rather than use extra space for the refcount). It's worth noting that most recent work on tracing GCs has been on reducing pause times, not improving throughput, so it wouldn't shock me if Immix is still the state of the art there.

There may have been further developments in improving RC throughput since 2013, but I'm not familiar with them; outside of Apple, there's very little modern research into optimizing reference counting. I know Swift does many cool optimizations of its own, but I'm not sure how restricted they are by having to remain compatible with Objective C.