Better results than the alternative, assuming your use case is to find two accounts by indexed ID, check a balance, and update two rows. And you have a small in-memory dataset, tiny rows, indexed point lookups, very little application computation, no external I/O, no joins, no analytical queries, no large scans, no disk-capacity pressure, thousands of independent requests available for pipelining, and deliberate hot-row contention. Or in other words, you have the perfect conditions for a...hashmap behind an RwLock!

Hm I wonder why people find this benchmark unfair...

You're just describing OLTP transactions, which is exactly what OLTP benchmarks are designed to test.