Path: utzoo!attcan!uunet!snorkelwacker.mit.edu!mit-eddie!wuarchive!sdd.hp.com!hplabs!hpfcso!mjs
From: mjs@hpfcso.HP.COM (Marc Sabatella)
Newsgroups: comp.sys.hp
Subject: Re: Re: Dream System
Message-ID: <7370270@hpfcso.HP.COM>
Date: 13 Dec 90 17:49:33 GMT
References: <54@gauss.mmlai.UUCP>
Organization: Hewlett-Packard, Fort Collins, CO, USA
Lines: 25

Another "fact" that needs clarification is the inferrence that somehow SPEC
numbers are more "meaningful" than MIPS numbers when comparing RISC vs. CISC.
This would be true only if MIPS really did mean "Millions of Instructions Per
Second".  However, no one actually calculates MIPS ratings that way any more.
They are calculated *exactly* the same way as SPEC numbers are - ie, you run
a few benchmarks, divide by the time the same benchmarks take on a VAX, and
produce some sort of average.  RISC/CISC is entirely orthogonal to this.

SPEC ratings differ from MIPS rating in three important respects.  The MIPS
benchmarks are almost entirely integer, almost entirely C, and almost entirely
"toys" (ie, Dhrystone).  The SPEC suite is more floating point intensive and
uses a lot of Fortran, and includes more "real" applications.  Neither suite
tells you much about overall system performance - except for gcc in SPEC, none
do much I/O, or use much VM; none attempt any graphics, networking, etc.

In theory, because of the way the numbers are generated, SPECmarks would always
equal MIPS ratings.  This is almost never true in practice, though - MIPS
numbers are artificially inflated because several of the benchmarks in the
suite are unrealistic "meatballs" for an optimizer to speed up.

--------------
Marc Sabatella (marc@hpmonk.fc.hp.com)
Disclaimers:
	2 + 2 = 3, for suitably small values of 2
	Bill and Dave may not always agree with me