Mainframs in the true sense nowadays are not massively powerful cpu based systems.
A mainframe is constructed of a series of highly optimised peripherals and core processing architecture all designed to be self sustaining.
So where we think of a cpu running disk io, printing, gui, application etc, the core of a mainframe just runs basic application calculation and OS. All the peripherals, disk, comms, i/o, etc. are relatively self sustaining components. This means that your disk drive itself is a pretty intelligent device that requires little control or maintanance by the central processor - it is all handled through local subroutines. This leaves the central cpu essentially free to do what it needs to - process data. It also means that uptime and stability are paramount, since the amount of central code running is so much drastically lower.
Mainframes are not good at complex 3D mappings, theoretical nuclear blast simulations etc. what they do very well is finance and banking. Stupendous levels of reliability and uptime, very good basic financial calculating ability and the capability to scale out to very large systems. They have phenominal running costs and incredible levels of QA and change control assosiated with them. Think of any component in a mainframe and you are usually talking 6 figures. An IBM FEP (Front Edge Processor - or user IO module to you and I) can be anything from $1,000,000+.
What is more glamourous - as is mentioned above - are the MPP (Massively Parallel Processing) machines and clusters, where 64+ cpus get together in 1 or more chassis and have a good old think about the world. These are the guys that simulate the big bang, nuclear explosions and such.
Hope this helps a little.
-* This Space For Rent *-
email for application details