All talk & no benchmarks. If reference counting with cycle detection was faster, the jvm & clr would use it.
When I cosnider tracing GC my primary arguments against it have nothing to do with performance at all, yet it often comes up as an argument against my position. This article is attempting to address that critique, by saying I doubt you could compare the two directly.
You "mention" this as if it were fact, yes. Apparently you are not aware of realistic comparisons that do exist, such as http://www.cs.utexas.edu/users/mckinley/papers/mmtk-sigmetri... (Alternatively, the article needs a discussion of why you think this is not a "realistic comparison".)
for a deeper discussion see: https://www.infoq.com/presentations/java-vs-c-performance