Author: Robert Hyatt
Date: 05:57:27 12/06/99
Go up one level in this thread
On December 05, 1999 at 23:00:03, Paul Petersson wrote: >On December 05, 1999 at 09:51:42, Robert Hyatt wrote: > >>On December 05, 1999 at 03:54:32, O. Veli wrote: >> >>> I would like to learn how much difference a dual-processor would make for a >>>chess program? Say a dual 500 Celeron and a dual PIII 450's. I assume it would >>>not mean a lot for much of the programs except for the Crafty and the upcoming >>>Deep Junior. How much difference could one except for these programs in Elo? >>>Thanks for your time. >> >>I would estimate the typical performance improvement (two cpus) at 1.6-1.75 >>times faster (regardless of estimates you might see that suggest better >>numbers). That is probably about 50 rating points better, since many quote the >>70 points per doubling of cpu speed... >> > >I´ve heard that ChessBase claims a speedup of 1.8 for Deep Junior, and that Deep >Junior is 5 percent slower than J6 on one cpu. > >Paul > No idea how they test. Bruce/I have been using what I would call a very 'unfriendly' test set... in my case one of the positions runs 10x _slower_ with 4 cpus than it does with 1. If you pick the right set of positions, I run 2x faster (or even far more in good cases). If you pick the wrong cases, I may run slower than 1 cpu. For a broad set including opening, middlegame and endgame, I see the numbers I gave. In a game, it probably does better since hash table stuff carries across searches and helps with move ordering.. > >>I have no idea what kind of speedup Junior gets. I have lots of good data for >>Crafty, that is in the speed-up range given above. Cray Blitz did better, but >>the algorithm was _much_ harder to write and debug, and using a recursive >>alpha/beta search I don't see any easy way to implement the DTS algorithm I >>used. >> >>Note that there are a few pathological positions that will tear up any parallel >>search. IE it is possible to run _slower_ with 2 cpus than with only one, on >>some odd positions. It is possible to run 10X _slower_ in such cases... Which >>is a headache...
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.