The variance is sometimes due to RVM troubles. The benchmark is partly a test of the entire Ruby ecosystem. If RVM is bugging, then people are installing sub-optimal implementations. For example: https://github.com/wayneeseguin/rvm/issues/1418