discrete transistors

From: William Donzelli <william_at_ans.net>
Date: Wed Oct 21 08:41:11 1998

> What amazes me is that no one has ever marketed a mainframe-like
> machine out of modern processors. Someone mentioned they had made
> one out of a 286 and some z80s, but why did no single company ever
> sell any? I would think such machines could be very useful.

One of the reasons (performance is another, but I think that has been
beaten to death) is fault tolerance. Most mainframes are installed in
banks and other financial institutions, where one day of downtime can be a
major disaster.

As far as I know, no x86 or 68xxx processor has ever had any degree of
fault tolerance withing the chip, where it is needed. Getting it external
to the chip generally involves triplicating everything and having voting
circuits, which can be a right royal, pain dealing with the error
conditions.

Most mainframes, even ones from the 1960s, have error checking throughout
the entire system - even in the paths between the registers and ALU. If
something goes wrong, like a gate goes into a "stuck at" condition, the
redundant circuits and error correctors will jump into action and
processing will not stop. Most machines will call home and have
replacement parts ordered and engineers busy. For example, a stadard joke
(quite true, too) amongst Stratus users is "how do you know when your
Stratus has broken? When the new part arrives the next day in FedEx."

William Donzelli
william_at_ans.net
Received on Wed Oct 21 1998 - 08:41:11 BST