Author: Vincent Diepeveen
Date: 05:42:55 04/25/01
Go up one level in this thread
On April 24, 2001 at 20:30:25, Robert Hyatt wrote: >On April 24, 2001 at 13:59:03, Vincent Diepeveen wrote: > >>On April 24, 2001 at 13:37:02, Robert Hyatt wrote: >> >>>pa4b5P Pa6b5p ra2a6 Nd6b7 bc5e3 Nb7d6 ra6a7 Bf8e7 >>> 9(6) #[axb5](49)#################################### 49 T=55 >>>pa4b5P Pa6b5p ra2a6 Nd6b7 bc5b6 Ra8a6r ra1a6R Nb7d6 >>>10(6) #[axb5](49)##################################### 49 T=160 >>>pa4b5P Pa6b5p ra2a6 Nd6b7 bc5f8B Qe8f8b ng3f5 Ra8a6r ra1a6R Rc8b8 >>>11(6) #[axb5](49)#[Nf5](50) 50 T=308 >>>ng3f5 Nd6f5n pe4f5N Pb5a4p bc2a4P Bd7a4b ra2a4B Qe8d7 bc5f8B Rc8f8b pf5f6 Qd7d5p >> >>Very confusing is whether it's a 11 ply PV or 12 ply pv. > > >What is the confuision. Hsu said "11(6) means 11 plies nominal in software, >6 more plies in the chess chips." What can _possibly_ be confusing about >that? Then Hsu is a big liar: <ch> 'c' --------------------------------------- --> 1. e4 <-- 39/119:51 --------------------------------------- Guessing c5 3(4)[Nf3](30) 30^ T=1 ng1f3 Qd8c7 3(5) 38 T=2 ng1f3 Nb8c6 4(5) 38 T=2 ng1f3 Nb8c6 5(5)[Nf3](52) 52 T=2 ng1f3 Nb8c6 6(5)[Nf3](68) 68 T=4 ng1f3 Nb8c6 nb1c3 7(5) #[Nf3](68) 68 T=5 ng1f3 Nb8c6 nb1c3 Ng8f6 8(6) #[Nf3](59) 59 T=6 ng1f3 Nb8c6 bf1b5 Qd8b6 nb1a3 9(6) #[Nf3](66) 66 T=8 ng1f3 Nb8c6 nb1c3 Ng8f6 bf1e2 Pd7d6 10(6) #[Nf3](53) 53 T=18 ng1f3 Nb8c6 bf1b5 Ng8f6 pe4e5 Nf6g4 bb5c6N Pd7c6b 11(6)<ch> 'e5' --------------------------------------- --> Pe7e5 <-- --------------------------------------- the 11 ply PV here doesn't show 11 ply or more. Knowing that at least 3 moves in this pv get extended it's very questionable whether they searched more as 5 ply in software here apart from the extensions. the 7(5) pv : ng1f3 Nb8c6 nb1c3 at the start of a game with 30 SP processors and probably sufficient hash, does that look like a 12 ply PV to you? To me it looks like a 2 ply PV + 1 pv extension and 5 ply in hardware... Don't tell me that every game they had bugs in hashing :) Best regards, Vincent > >> >>the moves with n or N behind it means captures. DB extends in software >>nearly all captures. I see around 3 non capturing moves here. > >Where do you get that? Their main extension is the singular extension. >Captures don't trigger that every time. > > >> >>So 5 or 6 ply in software + capture extensions >>(either recapture extensions or SE) >>+ 6 ply in hardware. >> >>Very logical. >> >>note it's a 12 ply PV you see here *not* a 11 ply pv. > >Vincent... you are trying to make up the rules as you go. I specifically >asked Hsu about the depth. Your interpretation is totally off-the-wall and >irrelevant to the discussion. He said that the 11 above means that the >software was started with a depth of "11", which would be reduced by one >for each ply searched, and extended as indicated based on their normal >extension logic. When depth reaches zero, they then give that position to >a chess processor which then searches from that position for N plies, where >N in this case is 6 more. The only difference is that singular extensions >don't happen in the hardware. > >The chess processors _can_ probe a hash table. But he reported that he did >not have time to design a 16-port memory module for each chess board... so >the 1997 version didn't do hash probes. But only because he didn't have time >to do it, not because it is impossible to do. > > > > > >> >>I get way longer lines at 12 ply with extensions turned on as this :) > > >So what. You don't have the same problem. They used 32 processors. A >single processor searched about 1/2 way thru the nominal software depth, >then positions were farmed out to 32 nodes on the SP (which use message >passing for communication and have distributed memory). After depth goes >to zero, the remainder of the search was done by chess processors that have >_no_ hash memory, and _no_ PV facilities... Which means they will _never_ >see the part of the PV searched by the chess hardware. And it is likely >that they will never see all of the software PV since the PV was searched >by different processors with no shared hash... > >Looking at the length of the PV is meaningless, and I have pointed this out >before. They can _not_ get a PV from the chess hardware. _period_. It is >simply impossible as it was not designed into the processors, and with no >hash table, there is no other way to do it... > > > > > >> >>>qf2g3 Pg7g5 >>>--------------------------------------- >>>--> 33. Nf5 <-- 7/65:41 >>>--------------------------------------- >> >>This is caused by 30 diff processors with SE implemented. >>I have those huge lines too in DIEP when i turn on all extensions! >> >>No big deal. >> >>If 6(6) would mean 6 ply in software and 6 ply in hardware, >>then why do we see only 5 ply line? > >See above. I get short PVs all the time. Since their hash is distributed >over 32 nodes, who knows what gets overwritten, and when... > > > > > >> >>Even if you overwrite on an SP computer you still get 6 ply! > >Nope. They reconstruct PV from hash, since they can't back the PV up >from endpoints due to hardware. > > > >> >>Now the theoretic impossibility of searching 17 ply fullwidth >>*with* all those extensions the first 11 ply. >> >>Apart that each search line must be like 15 ply then or so, >>It's going to use up a lot of nodes. >> >>For deep blue it would cost around 5^6 more as the nodes they got in 1997! >> >>>Again I reiterate, the notation 11(6) means 11 plies in software search, >>>6 plies in the hardware, plus the quiescence in hardware. There is _no_ >>>argument with this. Simply ask any of the DB guys. 11(6) is a total of >>>17 plies of search. >> >>Noop it is not Bob. It is 11 or 12 plies of search from which 6 ply >>in hardware. Makes sense. Logical and clearly visible from the lines. >> > > >No it doesn't. You don't get to directly contradict the DB team, just because >you can't imagine how it could be done. If half of the search were done by >the chess processors, you would _never_ see a 12 ply PV, since the chess >processors can't provide _any_ PV at all. > > > > >>The first few ply >> >>Note that if it would be 11 ply of search with pruning + 6 ply in hardware, >>then deep blue is the tactical worst program in history as it sees >>Bf5 in game 6 at 8(6) which would be 14 ply then, which doesn't make >>sense! Not even if you forward prune a lot! >> >>Shredder needs 8 ply for it too. It's a tactical queen win. >>Shredder is doing recapture extensions as far as i know. >> >>If i turn on extensions in diep i also need 8 plies. without recapture >>extensions i need 9 or 10 ply. >> >>Idem for other progs! >> >>Best regards, >>Vincent >> > >So? Db is not "other progs"...
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.