Such a claim really needs hard data to back it up. Reference counting can be very expensive, especially if the refcount update is an atomic operation. It's hard to capture in profiling tools because the performance overhead is smeared all over the code base instead of centralized in a few hot spots, so most of the time you don't actually know how much performance you're losing because of refcounting overhead.
The most performant approach is still manual memory management with specialized allocators tuned for specific situations, and then still only use memory allocation when actually needed.