Author: Christophe Theron
Date: 11:45:17 12/09/99
Go up one level in this thread
On December 09, 1999 at 12:39:08, Vincent Diepeveen wrote: >On December 09, 1999 at 11:57:17, Christophe Theron wrote: > >>On December 09, 1999 at 08:45:06, Vincent Diepeveen wrote: >> >>>On December 09, 1999 at 05:58:45, Enrique Irazoqui wrote: >>> >>>>On December 09, 1999 at 04:49:45, Enrique Irazoqui wrote: >>>> >>>>>The best I can do is to play 40 games at 60/60 >>>> >>>>I mean 60/30, 60 moves in 30 minutes. >>> >>>no way i'm never gonna go 60/30 then diep gets sometimes only 7 plies >>>deep in a few crucial positions. >> >> >>Having 6 times more time will not save you in this case. How deeper can you go >>with 6 times more time? 2 plies I guess? > >minimum depth 9 indeed. > >there is a HUGE difference between all plies *at least* 9 ply >or *at least* 7 ply. > >At duals and quads the bigger hashtable and parallel cutoffs >cause simply that the worst case is less bad of DIEP, as i don't >allow to split below my trees more than 0.5 n processors, >so there is no worst case there. > >This causes in positions where the single cpu version has a bit of >bad luck with move ordering, that the parallel version usual doesn't >have it. > >therefore at a quad xeon 400Mhz (linux), which represents 4x400 x 0.9 >(linux compiler slower) = 1600 x 0.9 = 1400Mhz roughly. > >that sounds a lot, but at 90 0 in dutch open i got all games at least >10 ply deep, where at home some positions at a 450Mhz NT at analysis level >it sometimes hardly gets 9, though simple math tells me that objectively >1400 / 450 = roughly 3.5 > >So if i give it 3.5 times more time than the quad got in dutch open, >then i should get objectively of course single cpu at least the same >depth in *any* position. > >I'm not interested in the speedup then of course, i'm only interested >in the minimum depth my program searches in this case. > >However this test goes wrong. the quad does *a lot* better. >Usual my move ordering does ok, but in a few positions where it is dead >wrong, there parallellism in my opinion clearly prevents that >worst case to happen. > >40 in 2 level at at least k6-2 (preferably intel though) >like at SSDF i find therefore a normal time control for this bet. > >Vincent > >>You need a 1 move in 2 hours time control. > >Such a match would be fun too, but a bit hard to get systemtime for. >i want to have my champaign sooner than that too... > > > >> >> >> >>>don't worry i've got offers enuf from testers. >> >> >>Who has offered to play the match? >> >> >> >> Christophe So now we just need a volunteer... 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.