This is a really good question. I executed the following two queries in the Northwind database on SQL Server 2000:

select *
from orders
where orderdate = (
select max(orderdate) from orders)

select *
from orders
where orderdate = (
select top 1 orderdate from orders order by orderdate desc)


The first query executed with a subtree cost of 0.00741. The second query had a subtree cost of 0.0190. Based on this, the first query is faster.

The first query produced the following query plan:

  |--Top(1)
       |--Bookmark Lookup(BOOKMARK:([Bmk1003]), OBJECT:([Northwind].[dbo].[Orders]))
            |--Index Seek(OBJECT:([Northwind].[dbo].[Orders].[OrderDate]), SEEK:([Orders].[OrderDate] IsNotNull) ORDERED BACKWARD)

The second query produced this plan:

 |--Bookmark Lookup(BOOKMARK:([Bmk1000]), OBJECT:([Northwind].[dbo].[Orders]))
       |--Nested Loops(Inner Join, OUTER REFERENCES:([Orders].[OrderDate]))
            |--Top(1)
            |    |--Index Scan(OBJECT:([Northwind].[dbo].[Orders].[OrderDate]), ORDERED BACKWARD)
            |--Index Seek(OBJECT:([Northwind].[dbo].[Orders].[OrderDate]), SEEK:([Orders].[OrderDate]=[Orders].[OrderDate]) ORDERED FORWARD)

So it looks like the optimizer has been trained to look for the first query and provided better performance. I'd stick with the first form.

-Mike