Author: Robert Hyatt
Date: 21:40:24 05/16/04
Go up one level in this thread
On May 17, 2004 at 00:19:00, Tom Likens wrote: >On May 16, 2004 at 20:57:34, Robert Hyatt wrote: > >>On May 16, 2004 at 19:10:30, Tom Likens wrote: >> >>>On May 16, 2004 at 16:54:10, Robert Hyatt wrote: >>> >>>>On May 16, 2004 at 15:39:18, Anthony Cozzie wrote: >>>> >>>>>On May 16, 2004 at 13:56:23, Tom Likens wrote: >>>>> >>>>>>On May 16, 2004 at 13:17:17, Anthony Cozzie wrote: >>>>>> >>>>>>>On May 16, 2004 at 11:10:01, Omid David Tabibi wrote: >>>>>>> >>>>>>>>On May 16, 2004 at 11:05:36, Anthony Cozzie wrote: >>>>>>>> >>>>>>>>>On May 16, 2004 at 10:14:07, K. Burcham wrote: >>>>>>>>> >>>>>>>>>> >>>>>>>>>>Has anyone here posted using dual 248s? >>>>>>>>>>wonder how 2 x 248 AMD compares to dual Xeon 3.2 with 1 meg cache running a 32 >>>>>>>>>>bit commercial chess program? >>>>>>>>>> >>>>>>>>>>http://www.newegg.com/app/viewProductDesc.asp?description=19-103-433&depa=0 >>>>>>>>>> >>>>>>>>>>http://www.monarchcomputer.com/Merchant2/merchant.mv?Screen=PROD&Store_Code=M&Product_Code=120140 >>>>>>>>>> >>>>>>>>>>kburcham >>>>>>>>> >>>>>>>>>Like Mike, I have dual 246s. Opteron is a great CPU, but the 64-bit software >>>>>>>>>isn't *quite* ready, at least on linux (imo). >>>>>>>> >>>>>>>>Running 32 bit chess engines on that system, how much speedup do you see in >>>>>>>>comparison to the fastest 32 bit dual system? >>>>>>> >>>>>>>I wouldn't know. I only run 64 bit linux :) It works pretty well, but there >>>>>>>are definitely a few bugs left. I'd say in another 3 months my system will be >>>>>>>good enough for me (new nvidia drivers, a few more kernel versions). >>>>>>> >>>>>>>anthony >>>>>>> >>>>>>>>> >>>>>>>>>anthony >>>>>> >>>>>>Anthony, >>>>>> >>>>>>Are you running 64-bit SUSE or one of the Red Hat flavors? I've got an FX-51 >>>>>>that I initially loaded 32-bit SUSE on (this was before SUSE supported SATA >>>>>>drives right out of the box) and was pleasantly surprised at how fast the >>>>>>32-bit programs ran (chess engines included). When I finally installed the >>>>>>64-bit version, I was unpleasantly surprised at how *slow* the 32-bit software >>>>>>ran (including and especially the various Linux engines I test against). >>>>>> >>>>>>I got the requiste 64-bit boost when I converted Djinn over to a true 64-bit >>>>>>program but was somewhat dismayed with the 32-bit slowdown of its sparring >>>>>>partners. I *could* dual-boot into a 32-bit version of Linux for testing but >>>>>>frankly that offends my sensibilities and seems like a bit of a waste. >>>>>> >>>>>>regards, >>>>>>--tom >>>>> >>>>>Hmm, I haven't tried. I usually test against crafty, and I compiled a version >>>>>in 64-bit mode. I would have thought that 32 bit apps would run quickly because >>>>>this is essentially _hardware_ emulation, but maybe not . . . . >>>>> >>>>>anthony >>>> >>>> >>>>I don't see his problem. 32 bit code runs the same on a 32 bit or 64 bit >>>>operating system. The opteron just doesn't get to use all its "stuff" in 32 bit >>>>mode. I (and others) have done this several times and didn't see a 32 bit >>>>program run slower on a 64 bit O/S... >>>> >>>>Not sure what is going on there... >>> >>>Hey Bob, >>> >>>Unfortunately, I can't recreate the numbers now since I got rid of the >>>32-bit version of SUSE on the FX. From memory what I saw was that if I >>>compiled my program under 32-bit SUSE, using the Intel 7.1 compiler and >>>profile-guided optimizations, I got about 1.1M nps. So far, so good, but >>>when I switched over to the 64-bit version and compiled the exact same >>>program (in the same manner) I could only get a top speed of about 450k nps. >>> >>>I didn't really worry about it too much since I was converting the program >>>over to 64-bits (which gave me all the speed back and a bit more). My >>>assumption was that it was a problem with the 32-bit version of the >>>libraries, but now I'm not so sure (especially, if I'm the only one seeing >>>the problem). I may try a couple of experiments tonight to either verify or >>>invalidate the old results. >>> >>>regards, >>>--tom >> >> >> >>Really makes no sense. The only thing different is that the O/S has to >>save/restore extra stuff when context switching (extra registers, etc). I have >>not tried your experiment however. I ran Suse-64 on the quad opteron I used, >>and I compiled for 32 bit or 64 bit by simply telling gcc which architecture to >>produce object code for. I didn't find any difference on normal 32 bit stuff, >>although Crafty ran significantly faster compiled for 64 bit, however. AMD is >>putting together a demo to show the difference as they market the 64 bit >>performance... > >You may well be right. I ran some simple tests this evening and the 32-bit >executable of the program was pretty fast (700K+ nps) which was faster than >I remember. This is especially significant since the changes I've been >making lately have slowed the NPS down somewhat, so this number is not that >far off the 64-bit version. It could be (i.e. it's likely) that I had >something incorrectly configured initially. I've also patched the 64-bit >version of SUSE (to the tune of about 208MB using the "9.0-PatchCD.iso" >image off their website) since installing it to correct for the SATA drives. >So unfortunately, at the end of the day it's probably not an apples-to-apples >comparision. Anyway, my moneys on operator error ;) > >BTW, are you still running 64-bit SUSE or did you switch over to Red Hat? > >regards, >--tom The 4-way box was AMD's. They were running 64 bit Suse. They have a "relationship" with them but I am not sure how much I can say about it as I don't know what is public knowledge... We will have a bunch of dual opterons soon. We'll have to make that choice but I suspect Suse...
This page took 0 seconds to execute
Last modified: Thu, 15 Apr 21 08:11:13 -0700
Current Computer Chess Club Forums at Talkchess. This site by Sean Mintz.