Computer Chess Club Archives


Search

Terms

Messages

Subject: Re: DB NPS (anyone know the position used)?

Author: Robert Hyatt

Date: 20:50:51 01/25/00

Go up one level in this thread


On January 25, 2000 at 21:21:53, Chris Carson wrote:

>On January 25, 2000 at 20:56:21, Peter W. Gillgasch wrote:
>
>>On January 25, 2000 at 08:46:27, Chris Carson wrote:
>>
>>>Does anyone know what position(s) HSU used
>>>to get the 100M NPS (DB) or 200M NPS for
>>>DBII?
>>
>>Why do you think that the position matters
>>at all ? As long as he can keep the chips
>>busy the total number of cycles should be
>>constant. Since he was doing the last 4 plies
>>in hardware - and that means that he basically
>>did all the positions in hardware - I suspect
>>that the overhead onthe SP which certainly
>>depends on the position can be neglected...
>>Obvious, isnīt it ?
>>
>>-- Peter
>>
>>>Was it one position or many position's?  Was it
>>>middle game, endgame, or combination?
>>>
>>>I would like to have the EPD for the
>>>position, it would be interesting to
>>>benchmark against.  :)
>>>
>>>Thanks.
>>>
>>>Best Regards,
>>>Chris Carson
>
>OK, but that is not why I wanted the position.  HSU
>quoted the numbers and I would like to know what positions
>were used.
>
>Best Regards,
>Chris Carson


I don't think _anyone_ uses positions when they quote nodes per second.  IE
when I quote 800K for crafty on my quad xeon, that comes from watching a bunch
of games and noting the NPS as the games are played.  It runs from as low as
550K to about 1M on my xeon.  But once castled, it averages about 800K for the
rest of the game...

NPS is a 'vague' number since I have positions where I can hit 450K or 1.3M
depending on which one I use.  Most of use use actual game numbers as a result,
although it doesn't mean a thing.

In Hsu's case, he can't _really_ give an exact nps, because the hardware
processors don't count nodes. It would take them as long to count one as it
does to search one.  He seems to have an idea that they drove the chess
processors at 50-70% utilization.  Which would translate into roughly 500-700M
nodes per second, raw numbers.  He also claimed 30% search efficiency, which
most likely turns into the 200M number.  IE for Crafty, the raw NPS is 800K.
But due to extra work done, this is probably equivalent to 600K overall.

I am not certain how he arrived at 200M, but that is a reasonable guess.  Based
on 480 chess processors at 2M to 2.4M nodes per second.  taking the lower number
(some were clocked at 20mhz, others at 24mhz) we get almost 1B nodes per second
max, but he couldn't keep them all busy.  .70 * 1B = 700M.  .30 * 700M = 210M,
which was the claimed speed...




This page took 0.11 seconds to execute

Last modified: Thu, 07 Jul 11 08:48:38 -0700

Current Computer Chess Club Forums at Talkchess. This site by Sean Mintz.