What is the expected behaviour of an atomic_ref with work_item scope? Does it make sense to apply a per-thread atomic?

The memory ordering enabled by the atomic_ref may still provide some benefit to individual threads, however, couldn't the same be achieved with atomic_fences instead of using an atomic_ref?

Are there any cases where work_item scoped atomics are meaningful, other than the aforementioned case?

How should an implementation handle work_item scoped atomics?