Can you post the actual execution plan for both queries here?

A wild guess is that parallell execution with hash match or sort is involved, and that the distribution of processed rows between each thread is so skewed that some threads run out of memory to store the hash buckets and therefore spilling them to tempdb.

Look for parallellism operators in the actual exec plan, and see if number of rows processed by each thread is very skewed. Also run a profiler trace on your spid with Sort Warnings and Hash Warnings included when executing the query.