Date: Sat, 11 Sep 1999 14:15:09 -0700 (PDT) From: Ross Harvey <ross@teraflop.com> To: don@calis.blacksun.org, sthaug@nethelp.no Cc: alpha@FreeBSD.ORG, rdabney@lasg.com Subject: Re: AXP pci/33 & memory question Message-ID: <199909112115.OAA02156@random.teraflop.com>
next in thread | raw e-mail | index | archive | help
> From: Don <don@calis.blacksun.org> ::: > > > [ axppci33 is slow ] ::: > > If you think a 233 is slow, you should try to the 100 Mhz 21064 :-) > > (I have it running here with -CURRENT, but it's impressively slow.) ::: > 233 slow? I have to wonder exactly what everyone here is trying to do with > their alphas? I use my 3 alphas as primary nameserver, secondary name > server and firewall. They are 21064 233's and they are extremely useable, > quite fast and in general a nice machine to work with. If you want to > greatly improve the speed of your alpha you really need to make sure it > has enough memory. 96 megs is the minimum I would put (and is what I do > have in mine) in for a useable system. Obviously more is always fun. These > are not the latest and greatest machine that Compaq makes and for how old > they are and how cheap mine were I think they are excellent machines. You guys are comparing apples and oranges, and getting confused. The axppci33 is a 21066, or `lca'. It's dog slow at any clock rate. The 21064, or `ev4', although it has about the same execution core, and although it came out _earlier_, is just plain way faster at the same clock rate, apparently because of its greater memory or cache bandwidth. Here is a simple benchmark I ran on several NetBSD/alpha systems: CPU System Time Cycles/S Cycles 21264 264dp 2.8 500 MHz 1400 M 21164 eb164 9.3 266 MHz 2474 M 21064 pc64 13.9 274 MHz 3809 M 21064 as200 17.4 233 MHz 4054 M 21066 Multia 53.4 166 MHz 8864 M A Multia is basically an axppci33 with lots of integrated peripherals and snazzy packaging. It is basically the same speed, adjusting for clock rate, but look out: _some_ axppci33's have no bcache, and the last thing a dog-slow processor needs is to run without a cache level. Anyway, note that the Multia requires 2.19x as many total cycles as my as200, which _also_ has a narrow dram memory bus (so the memory bandwidth folklore isn't completely correct) and 2.33x as many total cycles as my pc64. Also, for those who thought it was a reasonable test to race an old 164sx board with a previous-generation CPU running one compiler against a recent PII with a _different_ compiler, please note carefully the speed of a comparable (i.e., recent) alpha. Ross Harvey ross@netbsd.org To Unsubscribe: send mail to majordomo@FreeBSD.org with "unsubscribe freebsd-alpha" in the body of the message
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?199909112115.OAA02156>