ORM (Object Relational Mapper) is a tool that creates layer between your application and data source and returns you the relational objects instead of
(in terms of c# that you are using) ADO.NET objects. This is basic thing that every ORM does.
To do this, ORMs generally execute the query and map the returned
DataReader to the POCO class. Dapper is limited up to here.
To extend this further, some ORMs (also called “full ORM”) do much more things like generating query for you to make your application database independent, cache your data for future calls, manage unit of work for you and lot more. All these are good tools and adds value to ORM; but it comes with cost. Entity Framework falls in this class.
To generate the query, EF have to execute additional code. Cache improves the performance but managing the cache needs to execute additional code. Same is true for unit of work and any other add-on feature provided by EF. All this saves you writing additional code and EF pays the cost.
And the cost is performance. As Dapper does very basic job, it is faster; but you have to write more code. As EF does much more than that, it is (bit) slower; but you have to write less code.
So why your tests show opposite results?
Because the tests you are executing are not comparable.
Full ORMs have many good features as explained above; one of them is UnitOfWork. Tracking is one of the responsibilities of UoW. When the object is requested (SQL query) for first time, it causes round trip to database. This object is then saved in memory cache. Full ORM keeps track of changes done to this already loaded object(s). If same object is requested again (other SQL query in same UoW scope that include loaded object), they do not do database round trip. Instead, they return the object from memory cache instead. This way, considerable time is saved.
Dapper do not support this feature that causes it to perform slower in your tests.
But, this benefit is only applicable if same object(s) loaded multiple times. Also, if number of objects loaded in memory is too high, this will slow down the full ORM instead as then the time required to check the objects in memory will be higher. So again, this benefit depends on use-case.