Author: Robert Hyatt
Date: 19:20:27 02/02/02
Go up one level in this thread
On February 01, 2002 at 18:08:39, Ed Schröder wrote: >On February 01, 2002 at 15:10:14, Andrew Dados wrote: > >>On February 01, 2002 at 00:28:08, Robert Hyatt wrote: >> >>>On January 31, 2002 at 14:04:13, Andrew Dados wrote: >>> >>>> >>>> >>>>From their own publication, 'Deep Blue', June 2001 >>>>Example of search depths over one position >>>>r1r1q1k1/6p1/3b1p1p/1p1PpP2/1Pp5/2P4P/R1B2QP1/R5K1 w >>>>from DB-Kasparow game 2 from 1997, before move 37 >>>> >>>>When chips were set to minimum fullwith 4 plys: >>>> >>>>A.Iteration >>>>B.Minimum software depth >>>>C.Maximum software depth >>>>D.Maximum Estimated combined depth >>>> >>>>A B C D >>>>---------------- >>>>6 2 5 11-21 >>>>7 3 6 12-22 >>>>8 4 11 17-27 >>>>9 5 15 21-31 >>>>10 6 17 23-33 >>>>11 7 20 26-36 >>>>12 8 23 29-39 >>>> >>>>So iteration is clearly the sum of minimum software depth (B) and hardware depth >>>>(4 plys here). >>>> >>>>-Andrew- >>> >>> >>> >>>OK... but what does this have to do with the current discussion? DB doesn't >>>report "an iteration number". It reports things like 10(6) and directly >>>according to Hsu (from the email I posted) 10 is the software depth, and (6) >>>is the hardware depth. They are _added_ to get the total depth... >> >>Why would they publish a table to depth 12 if they searched till d=18 in real >>game? >> >>Recap: >> >>Arguments for depths of 17-18: >> >>1) Your email from Hsu >>2) DB logs, which show something, like 8(4) line followed by 8(6) line. >> >>Arguments against reaching d=18: >>1) Quotes by David Fotland from Dr Campbell on RGCC as I reposted here. >> >>2) According to their publication avg search speed over DB-Kasparov match was >>126M nps. As you and Ed noted ebf of DB is 4. No matter how they prune, those 2 >>numbers stand. >> >>Then time to finish depth 18 would be x*4^17/126Mnps, where x depends on search >>model, qsearch, extensions, SE etc. That x can not be less then 30 (no qsearch), >>more like 1000 for their search model. 4^17/126Mnps = 136 sec. >>for x=30 we get 68 minutes to finish depth 18; for x=1000 we'll get 2266 >>minutes. In the match DB searched for about 3 minutes/move. >> >>3) When DB sees some tactics in 10(6) line, is was noted that current PC >>programs see that in depths 10-12 (current programs heavily prune and extend way >>less comparing to DB). > > >>No matter what is true, you have to agree some things are not consistent here. > >Right. > >Now let's have a look at things from Bob's point of view and assume the >information is correct. Most of the time the logs shows 10(6) and 11(6). Can the >host (the IBM RS/6000 SP from 1997) do a 10-11 ply brute force search with all >those heavy extensions? If so, it then will all depend how fast the chess chips >are doing their 6 ply searches. Each chip is claimed to do 2-2½M NPS. I can not >find an average time for doing a typical 6 ply search in the hardware but if is >an accepatable time it is maybe doable? > >Ed > I tried several positions. I found one that had over 200K nodes at 6 plies. Others were much smaller (obviously simple endgames can be in the hundreds). at 200K, that is 100ms per search, pretty slow but with 480 processors, that shrinks quickly. For smaller trees, that goes down. I find it very hard to do any analysis with _my_ program that would in any way be equivalent to what they were doing. It is not trivial. If I had the hardware for a week, I could give some good numbers for what they were doing, of course. :) I can only assume what they said was possible, because they said they were doing it. Hsu said he could build a "belle on a chip" (belle was a box about 18" x 18" by 24" and most thought "right".) He delivered what he said he could. > >>-Andrew-
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.