> Batch inserts usually increase throughput by reducing the number of write IOPS, fsyncs, etc. They usually aren't associated with a 10x _CPU_ savings, which is the finding here.

Often the saving on roundtrips is bigger than any of these. Most database client libraries work synchronously, so if you insert via single row INSERT statements you'll approximately get a two context switches, and a roundtrip for each row. That's often more costly wall clock time wise than the insertion itself.