Author: Christophe Theron
Date: 14:51:10 03/27/01
Go up one level in this thread
On March 27, 2001 at 13:43:59, Vincent Diepeveen wrote:
>On March 27, 2001 at 11:24:44, Robert Hyatt wrote:
>
>The bottomline is that Christophe says there is no need to get
>a new parallel machine for a program as the speedup is from ratingpoint
>viewpoint worthless, sticking to the old machines is what he prefers to.
>
>Of course a great 'argument' to not make tiger parallel, something
>he'll never manage anyway.
I'm still wondering what you smoke usually.
Whatever it is, take my good advice and stop as soon as possible.
Christophe
>My argument is that if he buys a new parallel computer that
>the speedup is quite a bit more as on his old machine.
>
>Basically his redenation comes down to something like that you don't
>need to buy faster hardware at all as the extra gain from it
>is worthless.
>
>Now we can discuss long. But my point is that if i buy a new machine
>that it's usual searching plies more as the previous machine.
>
>Also there IS a big win over being parallel and being non-parallel.
>
>Whether it's 1 ply for someone or 2 ply for another that's no big
>deal here. the program gets more nps a second, just like
>a cpu in the future would do for it.
>
>Therefore i describe for people who are thinking of buying a new
>machine the big difference between my PII450 and my dual PIII800.
>
>If you don't find that a fair comparision, that's fine with me.
>
>But bottomline is that someone who used to say that only search
>depth is important (christophe) now says that the rating win
>of getting another ply is not so significant?
>
>I also remember a statement that a 300Mhz machine with tiger
>would beat a quad crafty easily on ICC.
>
>Bit of contradicting statements.
>
>For me it was already never obvious that searching deeper is
>going to play much better if your search depth is already quite
>big (say over 10 ply), but the lineair speedup of the program
>is out of question with new hardware!
>
>More processors means more speed simply.
>
>>On March 27, 2001 at 08:00:44, Vincent Diepeveen wrote:
>>
>>>On March 26, 2001 at 22:47:24, Robert Hyatt wrote:
>>>
>>>>On March 26, 2001 at 21:52:16, Vincent Diepeveen wrote:
>>>>
>>>>>On March 26, 2001 at 21:35:51, Christophe Theron wrote:
>>>>>
>>>>>>On March 26, 2001 at 20:06:21, Vincent Diepeveen wrote:
>>>>>>
>>>>>>>On March 26, 2001 at 18:49:47, Christophe Theron wrote:
>>>>>>>
>>>>>>>>On March 26, 2001 at 18:35:17, Frank Quisinsky wrote:
>>>>>>>>
>>>>>>>>>Hi Christophe,
>>>>>>>>>
>>>>>>>>>a very interesting and good message, like that !
>>>>>>>>>
>>>>>>>>>Now a little bit in my superb English :-))
>>>>>>>>>
>>>>>>>>>01. I believe you forget the important factor, that you can playing engine -
>>>>>>>>>engine matches on a dual system with ponder! An interesting player must not buy
>>>>>>>>>a second computer and an autoplayer.
>>>>>>>>
>>>>>>>>
>>>>>>>>I do not disagree, but my post was about the increase in playing strength from a
>>>>>>>>dual system. Maybe it was directed more at those who think that they are going
>>>>>>>>to crush everybody on the chess servers with an expensive dual system. :)
>>>>>>>
>>>>>>>I think those persons are not completely insane then, because
>>>>>>>in blitz programs search to very shallow depths. Especially the so called
>>>>>>>3 0 matches which are popular (i personally never play with diep 3 0
>>>>>>>against other engines as i don't care about that level as debugging
>>>>>>>such a game costs me more time as the game itself costs for sure, not
>>>>>>>to mention that the scores scroll too quick on my screen with 3 0 and
>>>>>>>the shallow depths make the game not very serious representing tournament
>>>>>>>level).
>>>>>>>
>>>>>>>Most people who are running a prog on a chess server kick only on dick
>>>>>>>height. In that respect a dual is a good thing for them. As in blitz it's
>>>>>>>not 20 points or so it's more like 150 to 200 rating points at least
>>>>>>>difference.
>>>>>>
>>>>>>
>>>>>>Your 150 to 200 elo figure is nonsense.
>>>>>
>>>>>I remember the first duals coming into ICC.
>>>>>Ratings jumped up. DIEP got completely annihilated by crafty
>>>>>when they arrived.
>>>>>
>>>>>Nonsense?
>>>>>
>>>>>Your words!
>>>>>
>>>>>Diep single cpu 450PII 6 ply in blitz.
>>>>>Diep dual 800 cpu 8-9 ply blitz.
>>>>>
>>>>>Pick the rating difference...
>>>>>
>>>>
>>>>
>>>
>>>PIII core is nearly 18% faster as a PII core for DIEP.
>>>
>>>More hashtables etcetera.
>>>
>>>The comparision is never fair of course. I had a 450PII 128mb RAM
>>>then bought a dual 256MB PIII800.
>>>
>>>That's how one must compare, *not* same Mhz versus same Mhz!
>>
>>
>>That simply produces gross errors based on different assumptions. IE if
>>you compare program X on a single CPU Y with program A on a multiple CPU B
>>machine, you have _lots_ of reasons why A is better than X. A is faster
>>due to the multiple cpus. A is faster due to the fact that CPU B is faster
>>than CPU Y. A is faster due to a faster bus on B than on Y.
>>
>>There is simply too many variables.
>>
>>I _always_ talk about parallel speedup as using one of cpu X, versus 2 of
>>cpu X. So that the _only_ variable is the number of CPUs. The mhz is
>>identical, as is cache, as is hash table size, etc.
>>
>>One degree of freedom lets you attribute an improvement to that one difference.
>>If you have multiple degrees of freedom you can't do that. Your program might
>>actually be worse on 2 cpus than on 1, but because you use 2 _better_ cpus,
>>it does better than it does on the poor 1 cpu platform.
>>
>>This is what correlation analysis is all about in statistics... and it has a
>>significant error rate for cases with > 1 degrees of freedom. With exactly 1
>>degree of freedom, the only error is in the actual sampling data itself, not in
>>attributing success or failure to one of several features that change together.
>>
>>
>>
>>>
>>>>I don't quite see how you get 2-3 more plies from a dual 800 over a single
>>>>450, nor do I understand why you compare the dual 800 to the single 450. On
>>>>a pure mhz basis, you are comparing 1600 to 450, which is almost a factor of
>>>>4. I can see 1+ plies with a factor of 4. I can't see 2-3 unless you are
>>>>talking simple endgames...
>>>>
>>>>
>>>>
>>>>
>>>>>>
>>>>>>
>>>>>> Christophe
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.