Ben Johnston eats the elephant:
One of the first things you learn when working with SQL Server, and other transactional based SQL systems, is that set based operations perform best. If you are querying data, a cursor pulling individual rows doesn’t perform as well as a single query. It doesn’t matter if that cursor is on the client side or the server side. A set-based operation is more efficient, runs faster, locks less, and is generally better than submitting multiple queries.
This is also generally true with delete statements. This post covers the exceptions to that rule. Large delete statements impacting many rows and large amounts of data (millions of rows and many gigs of data) can actually have decreased performance. With transactional systems, such as SQL Server, each transaction follows the ACID standard. Part of that standard ensures that transactional statements either complete or roll back fully – partial transactions are not allowed. For a delete statement, that means that all of the rows specified by the delete are removed from the table, or none are removed and the data rolls back to the original state. The delete and rollback behavior must be predictable and consistent or the data could be left in a contaminated, unreliable state. Performing very large deletes can present some challenges and needs to be treated differently in production systems.
Read on for the reasoning behind this, as well as several techniques you can use and how they compare.