Not a lot. In both cases it's a bunch of RPIs on a switch(s)!
If you don't have any software to support the message passing between processes on different nodes and organise job submission then it's all a bit pointless as they are all very much a bunch of separate computers on a network. This is where the Beowulf software comes in as it encapsulates the message passing PVM/MPI and the Job control. Large clusters often have entire suites of software and auditing for billing over and above this as well.
