Increase the tempdb data files by an equal amount to distribute the workload across all of the disks and files. Ideally, you want to have as many files as there are CPUs (taking into account the affinity).

I keep reading this over and over again for SQL Server 2005. Can anybody help me understand it? Shouldn't the number of files be determined by the number of drives, instead of CPUs?

I'm guessing it relates to the "multi-threading" capability of the server.

Well, yes, but as soon as a thread makes a blocking call (like disc or network IO) that thread is preempted and descheduled.