Newsgroups: comp.sys.transputer
From: jan@neuroinformatik.ruhr-uni-bochum.de (Jan Vorbrueggen)
Subject: Re: sneaky suspicion wanted
Organization: Institut fuer Neuroinformatik, Ruhr-Universitaet Bochum, Germany
Date: 16 Dec 94 12:03:41 GMT
Message-ID: <JAN.94Dec16130342@thalia.neuroinformatik.ruhr-uni-bochum.de>

In article <D0pn6C.LFA@info.swan.ac.uk> eeberger@faith (Lothar Berger) writes:

   PS: Subsequently my results for gnuplot. As you can see every forth result
   is smaller than the previous.

This is an effect of the instruction prefetcher. For all I know, and I have
done some intensive timing tests myself, you can rely on the cycle counts in
the manual, with the necessary caveats concerning memory accesses and the data
dependency of timing for fp ops of course.

	Jan



