Computer Chess Club Archives


Search

Terms

Messages

Subject: Re: multi-processor hardware question

Author: Aaron Gordon

Date: 13:56:27 06/03/03

Go up one level in this thread


On June 03, 2003 at 14:54:29, Robert Hyatt wrote:

>On June 03, 2003 at 13:18:15, Aaron Gordon wrote:
>
>>On June 03, 2003 at 11:44:02, Robert Hyatt wrote:
>>
>>>On June 02, 2003 at 09:39:27, Aaron Gordon wrote:
>>>
>>>>On June 02, 2003 at 00:34:31, Pavel Blokhine wrote:
>>>>
>>>>>On June 01, 2003 at 20:39:29, Aaron Gordon wrote:
>>>>>
>>>>>>On June 01, 2003 at 17:35:51, margolies,marc wrote:
>>>>>>
>>>>>>>hi all.
>>>>>>>as there are so many good deep engines out now...
>>>>>>>i am serioulsy considering assembling a dual processor setup for chess
>>>>>>>background analysis (on a home lan) and playing (competitively?)
>>>>>>>
>>>>>>>What I need from you guys are mobo recommendations uopn which to build a system.
>>>>>>>i have heard some of you write about an "iwill" brand board.
>>>>>>>of course I know the specs of the tyan and aopen boards already.
>>>>>>>Does someone here know if there are any opteron boards floating in the market
>>>>>>>place yet? ( or is this maximum overkill?)
>>>>>>>
>>>>>>>So, I'd appreciate any knowledge mainboard archeitecture recommendations.
>>>>>>>thanks- marc
>>>>>>
>>>>>>I'm not sure about motherboards now days for Athlon systems.. the iWill MPX2 was
>>>>>>the best one IMO. Now they don't make it any more, which is a shame because it
>>>>>>was awesome. About processors... I can take special Athlon XP 1700+ chips
>>>>>>(latest core) and modify them physically to run 2266MHz (2800+ is 2250MHz) on a
>>>>>>regular 133fsb motherboard and they'll run in SMP.
>>>>>>
>>>>>>There are no catches, you just pop in the processor and the motherboard will
>>>>>>detect it as an Athlon MP 2800+. You don't have to do anything special. The good
>>>>>>thing is they're extremely cheap AND faster than a dual Xeon 3.06. An added
>>>>>>bonus is you can pick up a dual Athlon board and two of these modified 2800+ MP
>>>>>>chips for less than the price of a single 3.06GHz Xeon.
>>>>>>
>>>>>>All chips will be fully tested to be completely stable. It's not really
>>>>>>necessary on the latst cores because all AMD is doing is dropping ~2.4GHz cores
>>>>>>onto an OPGA package and changing the 'settings' on the chip, marking it to
>>>>>>whatever the market demands, and selling it. All I'm doing is setting it back to
>>>>>>what it really is. :)
>>>>>>
>>>>>>Also, if requested, I could provide Athlon MP 3200+ to 3400+ chips (2.4-2.53GHz)
>>>>>>chips, these would technically be considered 'overclocked' and would run at a
>>>>>>slightly higher voltage than normal.. thus producing a bit of heat (still less
>>>>>>than a Xeon 3.06, however). I will still test for complete stability of course.
>>>>>
>>>>>
>>>>>Hello Aaron!
>>>>>
>>>>>Can you provide evidence to support your claim that AMD 2800 is faster than any
>>>>>Dual Xeon 3.06? I ask because one guy own a dual AMD 2800 at playchess.com and
>>>>>his Kn/s with Deep Fritz 7 and Shredder 7.04 and lower than mine in a dual Xeon
>>>>>3.06
>>>>
>>>>
>>>>I ran tests with Charles Worthington and without HT the Dual Xeon 3.06 he had
>>>>was slightly slower (slightly is probably an understatement) than a dual 2600+.
>>>>With HT it was a little slower (few percent) than a 2800+, we tested single cpu
>>>>HT & no HT, dual CPU HT & no HT. The engines we tried were Shredder 7, Deep
>>>>Fritz 7, Fritz 8 and Deep Junior 7. We didn't use any "fritzmark" methods as
>>>>I've seen those produce "odd" results. We tested nodes/second from a particular
>>>>position.
>>>>
>>>>Also, in crafty the 2800+ running an optimized binary is faster as well. Single
>>>>Xeon 3.06 got ~1.1 million kn/s, 2800+ got ~1.38 million. Take 1.1 * 1.8 and the
>>>>xeon should pull 1.98 million kn/s. 1.38 * 1.7 and the dual 2800+ should pull
>>>>2.346 million.
>>>
>>>
>>>Your math on the end is not quite right.
>>>
>>>The 1.7/1.8 numbers are _not_ NPS multipliers.  They are time-to-solution
>>>measures.  IE a program will find the move in 1/1.7th the time, even though
>>>its raw NPS will likely be almost 2.0 faster.
>>>
>>>At least for Crafty, (and excepting my dual 2.8 xeon which is the exception
>>>at the moment) my NPS will almost exactly double with two cpus, and almost
>>>exactly quadruple with four cpus.  But the time to find a particular move will
>>>not be 1/2 or 1/4, but more like 1/1.7 or 1/3.1 roughly).
>>>
>>>Don't depend on NPS to say much about parallel searching.  It is a _good_
>>>measure to compare hardware, of course.
>>
>>I understand completely, but in this case more nps is better.. since it'd be the
>>same program, same number of threads, etc. From the testing I've seen a dual
>>Athlon box with my binary gets a 1.70x speedup in NPS, dual Xeon gets 1.80x.
>>
>>Also, before I forget again..  ne question I had a while back was with Crafty in
>>the wild 7 position.. just reminded me about it. With some very brief testing in
>>linux w/ crafty being compiled under GCC I noticed the time to ply was much
>>higher with two cpus vs 1. I'll grab the latest crafty and see if it still does
>>it.. If so, any ideas as to whats going on? Will post back in a few with the
>>results of the latest Crafty on the Dual Celeron 550...
>
>Odd things happen in simple endgames, due to hashing.  The two processes
>interact by sharing info thru hashing.  And that can affect the time to
>solution significantly.  IE the time to a particular depth might be longer,
>but due to the shared hash info, the actual search might be "better"....
>
>That's what makes parallel search so much fun.

Indeed. Should be getting a dual Athlon board soon, going to try dual 2.4GHz
MP's (true 2.4GHz). Assuming the board can handle it perhaps we can have a "dual
duel" :)



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.