Author: Robert Hyatt
Date: 21:34:01 07/14/03
Go up one level in this thread
On July 14, 2003 at 16:38:44, Matthew Hull wrote: >On July 14, 2003 at 16:32:17, Robert Hyatt wrote: > >>On July 14, 2003 at 14:56:36, Matthew Hull wrote: >> >>>On July 14, 2003 at 14:27:12, Uri Blass wrote: >>> >>>>On July 14, 2003 at 13:38:30, Robert Hyatt wrote: >>>> >>>>>On July 14, 2003 at 03:27:27, Uri Blass wrote: >>>>> >>>>>>On July 14, 2003 at 00:00:53, Robert Hyatt wrote: >>>>>> >>>>>>>On July 13, 2003 at 15:03:38, Sune Fischer wrote: >>>>>>> >>>>>>>>On July 13, 2003 at 12:42:19, Robert Hyatt wrote: >>>>>>>>> >>>>>>>>>I (and many others) believe that the Elo system works well for players >>>>>>>>>that are pretty close in rating. It seems to work less well (in the case >>>>>>>>>of computers) for players that are significantly separated in ratings. >>>>>>>> >>>>>>>>I agree mostly, not sure why it should be different for computers though. >>>>>>> >>>>>>>I'm not either. But if you watch a 2000 computer play a 2600 computer, it >>>>>>>_seems_ to me that the 2000 computer wins more games than it should. Or at >>>>>>>least draws more than it should. I certainly can't prove this however, but >>>>>>>experience seems to (at least in my case) support this conclusion. >>>>>> >>>>>>What experience? >>>>> >>>>>On servers. >>>>> >>>>>At a couple of dozen ACM and WCCC and WMCCC events. >>>>> >>>>>on matches played here locally during testing. >>>>> >>>>>Etc. >>>>> >>>>>> >>>>>>If you use games on chess servers then it is possible that the 2000 computer >>>>>>simply updated the software but the result are still not written in the rating >>>>>>list so this is different experience than ssdf. >>>>>> >>>>>>If you are talking about static programs than based on my memory there was a >>>>>>version of cray blitz that beated Genius1 in every game. >>>>> >>>>>With a big hardware advantage. But It didn't win every game even though >>>>>it certainly should have. I don't remember the specifics now, but I played >>>>>something like 20 games and hit two or three draws. That was suggesting >>>>>a difference of 400+ rating points. The real difference was far greater. >>>>> >>>>> >>>>> >>>>>> >>>>>>Cray blitz had a big hardware difference but I do not think that the difference >>>>>>was more than 600 elo. >>>>> >>>>>At that point in time, we were talking about 500K nodes per second for >>>>>Cray Blitz vs genius on a 486/33, if I recall the hardware. The difference >>>>>was probably way more than 600 elo, based on human vs computer games against >>>>>both. >>>> >>>>It is possible that Genius has some weakness that humans could take advantage of >>>>it. >>>> >>>>Based on the ssdf rating list we have difference of less than 500 elo between >>>>Crafty(A1200) and Genius1(486/33 mhz). >>>> >>>>Crafty 18.12/CB 256MB Athlon 1200 MHz 2614 >>>>Chess Genius 1.0 486/33 MHz 2140 >>>> >>>>The difference is less than 500 ssdf elo and 500K nodes per second for cray >>>>blitz suggest that it was not better than Crafty on A1200 in the games that you >>>>played. >>> >>>Is this not the same bad NPS assumption that VD makes all the time? A CB node >>>is not a crafty node. >>> >>>Matt >> >>I mentioned that to Uri in my response to him. On a vector machine, you can do >>things that are impossibly expensive on a micro. It is not easy to vectorize a >>search, so the search doesn't go faster, but it is certainly possible to >>vectorize various parts of the engine so that they go like blazes, such as in >>move generation, attack detection, but most importantly, in static evaluation. >> >>Of course, you have to either know how to use vector hardware, or be willing >>to learn, neither of which fits Vincent and his many "impossible" comments. > >I'm guessing that the Motorola AltiVec (PPC G4) is not as nice a vector platform >as Cray, but could someone take advantage of it to measurable, though lesser >effect if they were so inclined? > >Thanks, >Matt > I'm not a good one to ask. I looked at this a while back, but only on paper to see what was included. It was really a "partial vector implementation" based on what a Cray does. It's more intended to just feed two operands to an instruction and get a result, and the repeat this N times. The cray does _much_ more than this.. IE You can vectorize a loop with an if in it... > >> >>> >>> >>>> >>>>I remember that latest Cray blitz could search 7M nodes per second but I >>>>understood also that you limited Cray blitz >>>>in the games against Genius1 so 500K nodes per second seems logical. >>>> >>>>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.