alpha anomaly report

[David Hough validgh business]

Stuart McDonald relayed the following from a recent issue of Infoworld:


        NOTES FROM THE FIELD by Robert X. Cringely(R)
        ...
        _Alpha waves_
        Elsewhere on Wall Street, a trader felt a
        tickle in his belly when he noticed that the
        price of a particular bond was showing up
        differently on two different computers,
        both Digital Alphas.  Fearing a bug in the
        most recent release of his software, he
        stepped through the code and discoved
        that the pow() function was returning different 
        answers on the two machines.
          Oddly enough, the machine with the
        newly revved motherboard was churning
        out errant numbers.  Test your Alpha thusly:  Try 
        pow(1.234567,7.654321).  If you
        don't get 5.017, call Digital tech support.
        ...
         


I have a 500MHz Alpha system intended for the http://www.validgh.com/linux-233
project, but I haven't unpacked it enough to check this out.   Anyway my
testing will be with linux, GCC, and libfdm, so I probably wouldn't encounter
the same anomaly even if it were in my hardware.   I have been
told that the 500 MHz Alpha chips used in NT systems are different from those
used in Digital Unix systems, in order to protect the profit margins on
the latter; I doubt that difference extends to the floating-point
instructions used to implement a pow function, however.

But the story illustrates an important point:   the money lost due to
numerical problems is not usually the direct cost due to bond miscalculations,
but the indirect cost of the time wasted tracking down what may or may not
have been an anomaly worth worrying about.

<PRE>