Author: Vincent Diepeveen
Date: 17:22:02 04/03/02
Go up one level in this thread
On April 03, 2002 at 15:00:05, Uri Blass wrote: I do not agree. 2 processors is for free getting nearly 2 times faster. If that requires effort, it is a peanut compared to the effort of writing a very efficient search. If you have a very efficient search, then again the argument is even harder: getting nearly 2 times faster is making your thing 2 times faster than the opponent. This removes simply disadvantages as being a slow searcher (in my case). I'm always faster than the guys who do not manage to get dual CPU. then let's talk about price. the fastest single cpu liquid cooled CPU is always slower for DIEP than DIEP at a dual CPU, whereas everyone understands that liquid cooled VAPOCHILL a look like stuff and the latest CPU is nearly impossible to get for me. Not to mention unaffordable (also hard to get is a chip that is fast AND overclocks well). In short being SMP removes always the 'i need to invest just too much money for this world champ' argument. If i show up tournaments with a dual 1.2Ghz K7 (which is pretty likely), then i could perhaps for the world champs upgrade it pretty cheap to 1.5Ghz dual K7 (above those speeds i already tried, the power supply doesn't give enough power for speeds above 1.5Ghz so i would need more money then to buy an expensive and better E-ATX PSU). Anyhow. Easier is just put in 1.4MPs pretty cheap upgrade. i'm nearly 3Ghz then (when talking about K7). a dual 1.4Ghz K7 is the same speed for DIEP like a P4 at about 6Ghz. So whatever you put single cpu against me, i'm faster simply. All those thousands of dollars you can invest for latest cpu. If you are dual you CAN be fast for little money (doesn't take away that usually you then also invest for expensive CPU's to be even faster, but that's other thing). We all know how hard it is nowadays to improve search. It is nearly impossible with a search to beat nullmove + efficient qsearch + nullwindow search (PVS). Of course you can shout about forward pruning, but before you get that to work you are 8 years further. Even then it is always doubtful it works at tournament level (considering you already get like 12 ply easily with just nullmove + well written qsearch). So let's assume that someone has put 8 years of his time in improving search, in fact most time is dedicated by most programmers in search improvements (amazingly i have to admit that even in DIEP half of my time still goes to search experiments). Suppose that someone manages to get a 5% node improvement by some kind of search improvement. He earns a nobel prize of course then if it is lossless (lossy is pretty risky and needs years of testing before it can be aproved working for him). Now compare that with the Huber algorithm which every good programmer who has a chessprogram can implement within a few days of his time. Another week debugging and he'll get easily a 1.6 speedup. How can you *ever* say that improving the search of the engine is smarter to do, considering that most programmers i know are basically busy with search and NOT other things. for an investment of at most 2 weeks and new hardware, you get nearly 2 times faster. I don't know a single thing that can speed you up that much! Of course you can go shout about hashtables and nullmove and pondering in the time of the opponents, but it is pretty idiotic to assume that someone doesn't use these known things. >On April 03, 2002 at 14:26:47, Nolan Denson wrote: >>Just wondering .. when writing a chess program why doesn't everyone program them >>to work in dual processing. Is it that much more difficult ... Also for some of >>the ones that are single engine now ... can you just simply take some common >>code from a program such as crafty and If so .. what parts of crafty does the >>dual processing ( lock.h etc etc)? It seems that many see writing there >>programs for dual processing is a lot of work. > >I doubt if it is a good idea to waste time about testing a program on 2 >processors when the advantage of it is clearly less than being twice faster. > >Things may be different if there are more than 2 processors and it is possible >to get a speed improvement of being 5 times faster but >I believe that today it is more productive not to work on teaching a program to >use 2 processors and to use the same time to improve the engine. > >Uri
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.