The current atomspace benchmark only measures node and link creation, and not 'typical' data. resulting in real-life performance that is one or two orders of magnitude worse than what's measured. See for example: opencog/test-datasets#7
The benchmark needs to be updated to measure more 'typical' data.