*BSD News Article 87730


Return to BSD News archive

Path: euryale.cc.adfa.oz.au!newshost.carno.net.au!harbinger.cc.monash.edu.au!lucy.swin.edu.au!news.rmit.EDU.AU!goanna.cs.rmit.edu.au!news.apana.org.au!cantor.edge.net.au!news.teragen.com.au!news.access.net.au!news.netspace.net.au!news.mel.connect.com.au!munnari.OZ.AU!comp.vuw.ac.nz!canterbury.ac.nz!newton!physmsa
From: physmsa@phys.canterbury.ac.nz (Mr M S Aitchison)
Newsgroups: comp.os.linux.networking,comp.unix.bsd.bsdi.misc,comp.unix.bsd.misc,comp.sys.sun.misc
Subject: Re: Sparc vs. x86 speed (was Re: Linux vs BSD)
Date: 29 Jan 1997 00:55:04 GMT
Organization: Physics and Astronomy Department, University of Canterbury.
Lines: 61
Distribution: world
Message-ID: <5cm758$d80$1@cantuc.canterbury.ac.nz>
References: <32DFFEAB.7704@usa.net> <5c58n9$hcb@innocence.interface-business.de> <32E66DE1.7E36AB48@samart.co.th> <5c8b0o$313$1@capsicum.wsrcc.com> <mumfordE4H6FI.Cr1@netcom.com> <32E75738.64D34A1F@tiac.net> <5cj8hj$1cm$1@capsicum.wsrcc.com> <32EDD877.52DB@Informatik.Uni-Bremen.DE>
NNTP-Posting-Host: newton.phys.canterbury.ac.nz
Xref: euryale.cc.adfa.oz.au comp.os.linux.networking:66492 comp.unix.bsd.bsdi.misc:5778 comp.unix.bsd.misc:2091 comp.sys.sun.misc:28126

In article <32EDD877.52DB@Informatik.Uni-Bremen.DE>, Michael Froehlich <mfr@Informatik.Uni-Bremen.DE> writes:
> Do you have a source for the $20,000 claim from above?
> *List* prices for the Ultra 1 starts at $8,000 (140MHz, 17", 
> 64MB, 2GB), street prices are lower...

Comparing systems is tough. Sun price low-end unexpandable boxes very low due
to market pressures (SGI etc do the same)... it is a big jump up to something
with a slightly higher clock speed and capable of taking more RAM, CPUs, etc.

This happens to a small extent in the PC market (a dual PPro motherboard is
much more expensive, but the percentage over the whole machine is nothing 
compared with the Sparc situation.  What does cost megabucks in the PC world
is a really good motherboard with better cache, etc. Given that memory bandwidth
is critical in a 200MHz computer coupled to 16MHz RAM you need something like
lots of cache and/or SDRAM (etc) to retain speed in large real-world jobs as
opposed to benchmark tests.  The UltraSPARC can take more cache than the PPro,
but it is one-way (as far as I recall), compared with PPros's two-way L2 cache
and even better 4-way cache on SuperSparcs (SuperSparcs were better than
hyperSPARCs and even sometimes UltraSPARCs if you run large progarms (e.g.
large matrices).  Floating point performance went up tremendously with both the
UltraSPARC and the PPro (compared with their respective predecessors), but both
are unimpressive compared with what has been available on IBM, SGI, HP and DEC
systems of a few years.  In fact the PowerPC 604e and soon the 704 are
much much faster (yet in the PPro price range), while the PowerIIsc (a bit more
expensive then a medium-sized Ultra) leaves everything else for dead. Assuming
we're talking about speed at certain types of CPU-intensive (especially
floating point) applications.

Often the best overall system speed is not just a matter of the raw CPU speed
but such things as the quality of compiler, the availability of two CPUs
(subjectively, our 2-CPU 55MHz SuperSparc outperforms many other systems
because there is a high change one of the CPUs will be available when you need
it).  And, if you can afford several computers, making one computer the "file
server" and another the "print server", etc can avoid expensive CPU cycles
being stolen from good compute servers that are easily (and reliably) put on
simpler systems.  With Intel boxes there is a much higher chance you can afford
several computers to fill the system's needs, which can provide backup
facilities (e.g. the file server can take over the print server's job at a
pinch). This might even mean you can get a good machine (dual 200MHz PPRo or a
modern PowerPC or a cheap Alpha) to take care of the compute grunt because it
now need not have the extra peripherals that push the price up.

I'd love to test a 300MHz ultra sparc; I have no idea whether the speed scales
well with the clock, but I know some things run well on a 200MHz Ultra
(compared with a PPro) and some things don't.  SPARCs in general are good at
context switching, and the availability of registers can help compilers produce
efficient code.  I get the feeling they are bad value, though. That's a
generalisation of course. If you *must* have the speed then you must be
prepeared to pay more at the high PPro end and above.  I find the price more 
than doubles when going to "reasonable" (i.e. a bit above bottom of the line) 
SPARC/SGI/whatever systems compared with the top-end PPro systems for what some
people find is worse performance.  You have to try *real* programs.  Sorry I
missed the original post - there might have been more details to go on.

-- 
-------------------------------------------------------------------------------
Mark Aitchison, Physics & Astronomy   \_  Phone : +64 3 3642-947 a.h. 3371-225
University of Canterbury,             </  Fax   : +64 3 3642-469  or  3642-999
Christchurch, New Zealand.           /)   E-mail: phys169@csc.canterbury.ac.nz
#include <disclaimer.std>           (/' "I am Daffy Duck of Borg. Resistance is dithpicable!"
-------------------------------------------------------------------------------