It's not a matter of "fractional speed difference" unless your database has very few entries. OR mismatch problems often like to appear shortly after your database starts to see any real use.
The only performant way to use an ORM is to use escape hatches everywhere. Alternatively, you can use an "ORM", something which calls itself an ORM while only doing superficial data mapping into dynamic or generated native (to your language) data structures. There are a _lot_ of these, most normal people call them query generators.
There are rather concrete problems that strictly prevent it from being possible to efficiently map graph (object) database access patterns to a relational database.
Do you mind going more into that? Naively, it seems like prolog/datalog describe graphs pretty well and they're inherently relational. Relational databases have typically just optimized for row-oriented OLTP uses instead of columnar OLAP, but there's nothing inherent preventing one or the other. They're duals of each other.