Author: Robert Hyatt
Date: 10:00:12 05/03/04
Go up one level in this thread
On May 03, 2004 at 12:19:36, Gian-Carlo Pascutto wrote: >On May 03, 2004 at 12:04:49, Anthony Cozzie wrote: > >>Honestly, I can't say that I understand the argument here. >> >>First, it is not clear to me that DTS implies any splitting method. So it >seems like, depending on your splitting method, your speedup could be 0-M, >>where M any number ;) > >That's correct - the papers I saw only give vague indications on where to >split. > >>Secondly, I didn't think Crafty used DTS. So it is not clear why results with >>Crafty reflect on DTS. > >Robert claimed nullmove makes no difference in speedup. I showed that >at least in Crafty, it does. 3% on one test. 10% on another. That does _not_ suggest that it makes a _BIG_ difference. That was Vincent's claim. I never said it made _no_ difference. I said "it makes no _significant_ difference." There _is_ a difference in the two statements if you read _carefully_. Vincent didn't need a 10% difference. He needed a 3000% difference or something. However he never posts any data so who knows how good or bad he really does... He quit posting after he fixed the bugs that made him produce > 2x speedup on every position he tried. > >I've seen no reliable data about DTS. I've seen reliable data nullmove >matters for speedup in a related algorithm. > >Given this, I see no reasons to believe Roberts claims about nullmove >not mattering anything until I see any evidence for them. At least the >opposite claim has *something* going for it. > >>In fact, I thought that almost no one was using DTS >>nowadays because it requires an iterative search. > >I don't know - it only makes much sense when you go multiprocessor, >and there aren't many people who've done that and talked about it. There has been one. And only one, so far as I can recall... > >>The only clear statement here seems to be that Crafty is less efficient with >>nullmove is on by 3-10%, depending on the position. Even if you prove the >>results are not identical (with X probability) it is clear null move is not >>_massively_ affecting the performance. > >Not for 4 cpus. I doubt that's going to be the area where the methods >distantiate themselves much from another in the first place. The hard part >is 16, 64, 512 cpus, and there this could matter A LOT. Who is talking about 512 except for Vincent? I have run on 16. DTS was run on 16. It was even run on 32 later after the T90 came out. I've _never_ mentioned going beyond that myself... > >-- >GCP
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.