Path: utzoo!attcan!uunet!cbmvax!bagate!rutgers!apple!agate!linus!linus!eachus From: eachus@linus.mitre.org (Robert I. Eachus) Newsgroups: comp.sys.amiga Subject: Re: A3000UX competition Message-ID: Date: 5 Dec 90 22:46:26 GMT References: <6057@crash.cts.com> <1990Dec5.061558.21893@engin.umich.edu> Sender: usenet@linus.mitre.org Organization: The Mitre Corporation, Bedford, MA Lines: 77 In-reply-to: gilgalad@caen.engin.umich.edu's message of 5 Dec 90 06:15:58 GMT In article <1990Dec5.061558.21893@engin.umich.edu> gilgalad@caen.engin.umich.edu (Ralph Seguin) writes: Preposterous. The 68040 at 25 MHz is MUCH faster than ANY 68030. I'd like to see the benchmarks and results you used. What benchmarks did you use? Also, please do not use pitiful benchmarks like dhrystones. Does anybody in the Amiga community have SPECmarks for them? I'd love to see what they are. I love the Amiga just as much as the next guy, but please do not post clearly false results. I will dig up some benchmark code from somewhere and run it on an 040 NeXT and then on a 50 MHz 030 Amiga and we'll see what we get. (FLAME ON!!!!) Come on Ralph, get your head on straight. If I were running dhrystones out of on chip cache, I would expect the 25MHz 68040 to be 50% faster than the 50MHz 68040. BUT I WAS BENCHMARKING SYSTEMS. Using "real" application code that required lots of memory accesses, in this case on a 1 Mbyte array. I like this particular program as a benchmark because 1) it is small (about 2K source lines), 2) it is very insensitive to compiler optimization, and 3) it has a good balance between integer operations, branching code, and memory references. The various Mac2's, including the 2fx, and the NeXT get dragged down by poor bus speeds. I wanted to check how good the GVP card was under stress. (Answer, good but expensive.) If you do lots of floating point operations the 040 will blow the 68882 and most RSIC chipsets away. BUT I WAS TALKING ABOUT INTEGER PERFORMANCE AND SAID SO. Different people use machines for different things. Typically, I need to benchmark for integer and byte manipulation operations, since the systems I deal are either used for discrete event simulation and message processing or have attached signal processors. (The Video Toaster is one hell of a signal processor, but not the kind I usually get asked about. :-) Last but not least: Don't insult Dhrystone, particularly Dhrystone 2.1. It is a very good and well designed benchmark. Just don't believe anything vendors tell you about Dhrystone numbers. When I was benchmarking a (name deleted to avoid singling out one of the guilty) in the manufacturers demo facility, I got a Dhrystone figure less than one third of what their marketing literature quoted. I got ahold of their performance person in the manufacturing plant, and told him my figure. I then "guessed" at 5 of the ways they had cheated to get the quoted number but that left me about 25% short. He told me that the marketing number used a timing package that excluded all system calls and part of the procedure call overhead! Yeech. It used to be, when the SPECmark first came out, the numbers were trustworty. I've recently seen some quoted numbers which were worse that Dhrystones. I've personally done thorough benchmarking jobs on over thirty systems. (And quick and dirty benchmarking of lots of others.) If you don't do your own benchmarking, and know all the gotchas, almost every marketdroid will try to pull considerable wool over your eyes. Commodore is one of the few that don't. (But then again they don't need to: The manual for my 2630 card: "...the A2630 may deliver a 400% - 600% speed increase in most integer based operations." The lowest improvement I have measured was 5.36x a base A2000 with FAST memory, the highest 10.7x. Great show Dave Haynie.) (Flame off.) I guess that got it out of my system. It's just that it seems all to often some procurement officer believes the marketdriods claims, and then I (or someone else) gets called in to study the "problem" with the software. It would be nice if the purchasers believed the reports of the people they hired to evaluate the systems instead of the vendors claims. -- Robert I. Eachus with STANDARD_DISCLAIMER; use STANDARD_DISCLAIMER; function MESSAGE (TEXT: in CLEVER_IDEAS) return BETTER_IDEAS is...