Author: Omid David Tabibi
Date: 06:17:31 04/12/03
Go up one level in this thread
Quite an interesting read Vincent. I'm afraid you are investing too much in the parallel speedup though. Any hardware speedup will be linear (at best) while algorithmic enhancements are exponential. If you manage to search one ply deeper by an algorithmic improvement, the gain will be more than any parallel speedup can yield. On April 11, 2003 at 08:34:50, Vincent Diepeveen wrote: >On April 11, 2003 at 08:28:55, Vincent Diepeveen wrote: > >note that they did internally count the number of searches done a second, >because a hardware node would get aborted within 8000 searches of all other >processors if it timed out. > >In short there must be a good reason they do not print it... > >>On April 10, 2003 at 17:33:49, Omid David Tabibi wrote: >> >>>On April 10, 2003 at 12:44:46, Vincent Diepeveen wrote: >>> >>>>On April 10, 2003 at 12:39:59, Christophe Theron wrote: >>>> >>>>>On April 10, 2003 at 12:02:05, Vincent Diepeveen wrote: >>>>> >>>>>>On April 10, 2003 at 11:37:50, Jonas Bylund wrote: >>>>>> >>>>>>>On April 10, 2003 at 10:27:57, Vincent Diepeveen wrote: >>>>>>> >>>>>>>>On April 10, 2003 at 10:11:21, Jonas Cohonas wrote: >>>>>>>> >>>>>>>>>On April 10, 2003 at 09:25:09, Vincent Diepeveen wrote: >>>>>>>>> >>>>>>>>>>On April 10, 2003 at 09:20:15, ERIQ wrote: >>>>>>>>>> >>>>>>>>>>>These are all great goals, but I like this order better. >>>>>>>>>>> >>>>>>>>>>> >>>>>>>>>>>* A Linux/bsd version of Chess Tiger w/ great gui ie. Eboard or better. >>>>>>>>>>>* A native ARM version of Chess Tiger for Palm >>>>>>>>>>>* Chess Tiger 16 >>>>>>>>>>>* ...and a few more projects that I prefer to keep secret >>>>>>>>>>> >>>>>>>>>>>basis for order is: >>>>>>>>>>> >>>>>>>>>>>1.I will have a freebsd system running in about two week hopefully (just waiting >>>>>>>>>>>on hardware to arrive) >>>>>>>>>>> >>>>>>>>>>>2.As soon as I could buy a new sony palm I will. So I can win a game from time >>>>>>>>>>>to time :) >>>>>>>>>>> >>>>>>>>>>>3.And ct16 should be last because ct15 is already too strong!! whether it's >>>>>>>>>>>first or last on that silly list that everone likes, I can't beat it on a 486 >>>>>>>>>>>comp. And yes I've tried shamlessly >>>>>>>>>>> >>>>>>>>>>>Just my two cents. >>>>>>>>>>> >>>>>>>>>>> Sign, >>>>>>>>>>> Eriq >>>>>>>>>> >>>>>>>>>>a dual version of CT15 would kick more butt than ct16 or working at the unknown >>>>>>>>>>freebsd. note that freebsd allows multiprocessing but multithreading at it i >>>>>>>>>>cannot advice. >>>>>>>>> >>>>>>>>>I think a dual version of CTX would be great! maybe we should have a hands up >>>>>>>>>here, to see if we can influence the order of things ;) (note: people with dual >>>>>>>>>processor systems votes count double, ok maybe only 1.7 :) >>>>>>>>> >>>>>>>>>Jonas >>>>>>>> >>>>>>>>Some people are simply too lazy or have too much bugs in their software to get >>>>>>>>stuff parallel well to work. >>>>>>>> >>>>>>>>It will be always like that. >>>>>>> >>>>>>>Well i don't know where that came from, but i am quite sure that an accomplished >>>>>>>programmer like CT would have no problems making Tiger SMP. >>>>>> >>>>>>in which case he is just plain lazy now. >>>>> >>>>> >>>>> >>>>>Vincent I have explained that I have other priorities. >>>>> >>>>>Why should I spend time on a task that I estimate is mostly a waste of time? >>>>> >>>>>Can you mention the benefits that has brought to you the multiprocesor version >>>>>of Diep? >>>> >>>>i show up at 500 processors world champs 2003 to name one. >>>> >>>>other is that i don't need to waste my time with forward pruning because i get >>>>already a factor 2 nearly out of SMP. >>> >>>You get a factor 2 speedup, so you don't need to waste time with forward >>>pruning?? Beware Vincent, you are starting to sound like the IBM folks...! >> >>in fact i have more processors than that 28 node 120Mhz (and 2 nodes 135Mhz) IBM >>machine had that ran deep blue hardware processors. Follow ups of this machine >>got sold then 'based upon deep blue technology', though i would swear to you the >>hard job was done by 480 hardware processors. It was NOT a shared memory machine >>at all, because all nodes communicated via a high speed switch with each other. >>So latency was poor to mention 1 problem of them with regards to parallel >>speedup. >> >>I run on a 1024 processor machine, fastest machine in europe to run chess >>software at and get 500 processors from that! >> >>In contradiction to deep blue the processors do not idle. >> >>Their speedup is based upon extrapolation of what the speedup is at a single >>node. They measured what a single node with single hardware processor did >>and compared that with what a single node did with 24 hardware processors. >> >>However the hard problem is the parallellization over nodes without shared >>memory. In fact all it had was a highspeed switch. Not even routers seemingly. >> >>To quote them more or less: "due to time constraints the nullmove search was not >>used" in deep blue hardware chips. >> >>I will not use that excuse 5 years after a match of diep versus some GM! >> >>Another werid thing is the number o fnodes a second. They cannot know it. their >>logfiles show no evidence of anything how many nodes a second they got. You can >>measure how many searches a second you do. They could know how much nodes a >>single search is on average. They cannot know how the 'big machines' number of >>nodes a second is. >> >>My estimate it 10 million nodes a second initially. Only in endgame perhaps 50 >>million nodes a second. >> >>We know SE was more limited implemented than i have done it. Also forward >>prunign was used in the hardware (no progress pruning for example). >> >>Their own estimation is a 8% speedup in tactical positions and 12% speedup in >>quiet positions based upon 'indirect evidence'. >> >>That is a joke estimation IMHO. Show me the number of searches performed first, >>because we all know the problems of an aphid type of parallel search. >> >>Best regards, >>Vincent >> >> >> >>> >>> >>>> >>>>>Have you won any tournament thanks to it? Has it improved your image, the >>>>>perception of quality in the eye of your future customers? How much money (=fuel >>>>>to continue improving your chess engine) have you made from it? >>>> >>>>It definitely will. >>>> >>>>Noomen has brought you a few victories at quick levels. Well done Christophe. >>>> >>>>>How can you justify that it has not been an almost complete waste of time? >>>>> Christophe >>>> >>>>History will forget you. History won't forget me. You will see soon. So far SMP >>>>was just a factor 1.5 speedup for most and 1.8 to 2.0 for some. >>>> >>>>But coming years the difference will be way more. Factors 4.0 to 8.0 will not >>>>get uncommon. Just read my lips. >>>> >>>>In my case how about 500 cpus at europes fastest supercomputer for >>>>computerchess? >>>> >>>>Now *that* is a bigger sales argument than you might think Christophe. Let's not >>>>publicly discuss sales. But you know what i mean. I bet you want to exchange all >>>>those tournament victories for just one shot at the title at a world >>>>championship which simultaneously is also a shot to playing the FIDE world >>>>champ! >>>> >>>>Best regards, >>>>Vincent
This page took 0.01 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.