Author: Matthew Hull
Date: 14:00:04 07/14/04
Go up one level in this thread
On July 14, 2004 at 16:53:39, Omid David Tabibi wrote: >On July 14, 2004 at 16:44:00, Peter Berger wrote: > >>On July 14, 2004 at 12:56:10, Omid David Tabibi wrote: >> >>>On July 14, 2004 at 12:34:18, Peter Berger wrote: >>> >>>>On July 14, 2004 at 12:28:29, Omid David Tabibi wrote: >>>> >>>>>On July 14, 2004 at 12:19:24, Peter Berger wrote: >>>>> >>>>>>On July 14, 2004 at 11:41:04, Omid David Tabibi wrote: >>>>>> >>>>>>>On July 14, 2004 at 11:38:31, Peter Berger wrote: >>>>>>> >>>>>>>>On July 14, 2004 at 11:26:47, Omid David Tabibi wrote: >>>>>>>> >>>>>>>>>On July 14, 2004 at 11:12:14, Fernando Villegas wrote: >>>>>>>>> >>>>>>>>>>Not at all, Omid >>>>>>>>>>If you already have a parallel engine you should run it into a hardware capable >>>>>>>>>>of getting all its power. >>>>>>>>> >>>>>>>>>I understand that you are going to provide the hardware, right? >>>>>>>>> >>>>>>>>>It is not something personal; next year I will have the needed hardware, but >>>>>>>>>what about others? Deep Sjeng and ParSOS were also parallel engines, but ran on >>>>>>>>>single processor not because they thought it was better, but because they did >>>>>>>>>not have access to a fast multiprocessor machine. >>>>>>>>> >>>>>>>> >>>>>>>>I wonder what except the results changed from two weeks ago to now to make you >>>>>>>>imply this is an unfair event and go on raving about it ?! >>>>>>>> >>>>>>>>Weren't you even one of the organizers? >>>>>>> >>>>>>>Yes, and I complained loudly about it even before the event. >>>>>> >>>>>>Actually I still think this is argueing with hindsight. Diep and Crafty finished >>>>>>3rd and 4th, a result that surprised some of the present programmers. >>>>>> >>>>>>It seemed to concern no one last year when it was only the professionals who >>>>>>came with superior hardware. >>>>>> >>>>>>To quote another programmer from memory: "I typically get 80% in tests against >>>>>>Crafty, so on a Quad it might get 30% or sth like that ". >>>>>> >>>>>>And we don't know how the results would have looked like in a single CPU event >>>>>>either - before the tournament it was "Crafty is a miserable program that I can >>>>>>beat on any hardware" (simplified and no quote) - >>>>> >>>>>This quote is not from me. >>>> >>>>And I didn't imply any different. >>>> >>>>> >>>>> >>>>>>at least one part of this >>>>>>statement turned out to be untrue. >>>>>> >>>>>>Maybe the hardware impact is a little overrated in this discussion anyway. Jonny >>>>>>on a PIV2.8 e.g. nearly finished Junior on a Quad. >>>>> >>>>>You want to measure the hardware impact, just do the following: >>>>> >>>>>Run the Crafty benchmark at a fast single processor machine, and also on the >>>>>quad machine: >>>>> >>>>>ratio = (benchmark on quad) / (benchmark on single proc) >>>>> >>>>>(the ratio would be about 4) >>>>> >>>>>Next, for each move of Crafty in the tournament, divide the thinking time by the >>>>>'ratio' above, to get a new time: >>>>> >>>>>time2 = (original time spent on the move) / ratio >>>>> >>>>>Now just count how many times Crafty changed its PV after 'time2'... (In other >>>>>words, how many times Crafty changed its PV after thinking for 1/4 of the time >>>>>on the move.) >>>> >>>>This test sounds more scientific and logical than it is. You also would have to >>>>identify the moves that would make any difference to the result of the game. >>>> >>>>I can easily reproduce all critical moves of the Falcon-Crafty game on my >>>>notebook for black. I already asked you in another message if more time would >>>>have made any difference for Falcon for two of its critical mistakes - you >>>>didn't answer yet. >>> >>>Just count the number of times Crafty changed the PV after 1/4 of its thinking >>>time (including pondering), and see how many of them were critical, and >>>prevented a loss. >>> >>>You will find many of them. Here is one I remember clearly: After Falcon's >>>...Ra5 in the endgame, Crafty wanted to repeat the position at first (score: >>>0.00). It thought for a long time, and then changed the PV to other move, which >>>was winning. But you will find many other moves in the earlier stage of the >>>game, which would have resulted in a loss were they played (after thinking for >>>1/4 of the time). >> >>I think you maybe misunderstood me here because of the heat of the discussion. >>That I can easily reproduce Crafty's moves on my notebook doesn't mean that all >>its moves were found fast and easily :). >> >>From memory there were two moves in the tournament where Crafty changed its mind >>very late and that turned out to be of big importance. >> >>Kb6 (or whatever the move after Falcon's Ra5, no surprise that we both remember >>this one) > >OK, so at least you agree that without the 4x hardware advantage Crafty wouldn't >have got more than a draw. > > >> and g5 in the game against Diep. >> >>In case there were other moves like that, I am at least not aware of it. >> >>I expect Bob to offer Crafty's logfiles for download soon as he did for cct-6, >>so maybe we should delay this discussion until we can just discuss the real data >>( I don't have access to the Opteron anymore). >> >>I would still be interested to know about the critical points from Falcon's >>side; Ra5 is a trial to avoid repetition - you have to search very deep to >>dismiss it. > >One iteration deeper Falcon switches to the other move which results in actual >threefold repetition. > >In other words, had Crafty not had the hardware advantage it would have missed >the win, and had Falcon had the 4x hardware it would have found the correct draw >(and with that hardware advantage probably also found a correct winning line). >Right? Why don't you just test it. Download crafty19.15 and setup the appropriate time controls for each side and start from some move in the actual game. Then we won't be in fantasy land anymore, but in the land of fact. :) > > > >>And Ra8 is probably not about search at all. >> >>Peter
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.