We did several benchmarks back in the day and it definitely depends on the size of your migrations + fixtures that you need in every test. Performance increases were simply huge for us, despite the serial execution of our tests in our legacy backends.

A new database scaffolded from a predefined template becomes available in several milliseconds or few seconds VS. truncating, migrating up and seeding (fixtures) would always take several seconds. We initially experimented with recreating these test-databases directly via sql dumps, then binary dumps until we finally realized that PostgreSQL templates are the best option in these testing scenarios.

Sorry, I've got no benchmarks to share at this time.