We used to maintain SGI boxes for this same reason but we've pulled back. The main benefit we got from older RISC stuff was that they bitched about unaligned loads. If you are going to support any of those models you need at least one box like that in your cluster. SPARC was faster than MIPS so we kept SPARC.

I think Theo could probably thin down the cluster and still be good but maybe I'm wrong and he'll show up with examples that require all of that hardware. That would be interesting because in our case we've sorted out the problems and rarely see things blow up on the RISC boxen.