Building on my previous post, the LRZ course slides show the following performance comparison:
Indeed one can see a two- to four-fold difference (and not 100!) as a result of memory layout issues linked to usage of derived types.
As @Beliavsky notes, the author of the ResearchGate post made other erroneous comments. Without concrete examples, it is impossible to say if and what went wrong in his case, or if it was simply an exaggeration.