Author: Robert Hyatt
Date: 12:37:48 08/29/01
Go up one level in this thread
On August 29, 2001 at 14:11:16, Uri Blass wrote: >On August 29, 2001 at 13:20:07, Robert Hyatt wrote: > >>On August 29, 2001 at 11:41:56, Uri Blass wrote: >> >>>On August 29, 2001 at 10:41:45, Robert Hyatt wrote: >>> >>>> >>>> >>>>If you _really_ believe that DF/DJ are better than DB of 1997, you are >>>>going to somehow have to overcome the following problems with that stand: >>>> >>>>1. We _know_ that DT played at a 2650+ level. Here is an excerpt from an >>>>email sent to me by Hsu: >>>>This is in the book draft that you have. The best 25-game performance was over >>>>2655. The USCF rating was much lower, because the first 20 games in DT's >>>>career, which include the disasters in early phase of the US OPen, had four >>>>times the weighting of the last 35 or so games in DT's career. This is >>>>so because DT's first official rating after 20 games was well over 2400, >>>>and the subsequent games therefore had only 1/4 weight. This is somewhat >>>>illogical, but DT was an unusual case to begin with. If all the games had >>>>the same weighting, then DT's USCF rating would have been over 2600, even with >>>>the disasters included. >>>> --FH >>>> >>>> >>>>So the 1992 (roughly) version of the program, made using 3 micron ASICS, >>>>was at _least_ as strong as either deep fritz or deep junior on today's >>>>hardware. I will be conservative and say they are "equal". >>>> >>>>We know that in 1992 they were running at around 1M-2M nodes per second, >>>>which is roughly what DF/DJ will do on an 8-way box. We also know their >>>>actual results against a bunch of GM players in multiple tournaments and >>>>matches hit a best of 2655 over _25_ consecutive games (not 25 cherry-picked >>>>games). >>>> >>>>2. The final DB machine was based on .6 micron ASICS. That is, each chip >>>>contains 25X as many logic circuits as the original 3.0 micron processors (if >>>>you reduce the width and length by 5, area is decreased by 25. So the final >>>>DB2 chips had 25X as many circuits. And we also know that 2/3 of the final >>>>chip was in evaluation. And being generous, 50% of the first chip was eval >>>>since he said "it is now up to 2/3 of the total chip area" which suggests that >>>>DT's chip was not nearly as sophisticated in the evaluation. >>>> >>>>Now let's do the math. The new DB2 chip has about 50X as much evaluation >>>>circuitry as did deep thought. The new DB2 machine was over 200X faster than >>>>deep thought. The original deep thought was at _least_ as strong as today's >>>>best programs. >>>> >>>>Conclusion: If DB 2 had 50X as much "smarts", plus was 200 times faster, do >>>>you _really_ conclude that DB2 is weaker than Frita or Junior? _really_? There >>>>is no way the math will support such a statement. >>>> >>>>A 200X faster DT would have been quite strong (that was basically what DB1 was >>>>all about in fact, more speed, but not a lot more smarts). A 200X faster, and >>>>a 50X smarter program sure seems, at least to me, to be significantly beyond >>>>anything running today on PC hardware or even on supercomputers. >>>> >>>>How do you counter those arguments, each of which is technically accurate and >>>>given by Hsu's publications or email that I have supplied. >>>> >>>>To say that DF/DJ is stronger is absolutely nonsensical... with nothing to >>>>support such a conclusion whatsoever... >>> >>>I believe that the rating of 1992 cannot be compared >>>with the rating of today so we cannot use the results >>>of deep thought against humans to learn that deep thought was at the >>>same level that Fritz is today. >>> >> >>What is this based on? I would personally bet that 2655 in 1992 is >>better than 2655 today, with the inflation in ratings that has occurred. >>However, I left that out of the equation so that DT's 1992 rating would >>look as weak as possible when compared to today. > >2655 in 1992 is better place in the world than 2655 today but it does >not prove that it is better level of play. the pool of players is not that different. So I simply say that the two are "close" if not equal and move on... > >> >> >> >>>I believe that today humans know better how to play against computers >>>relative to their knowledge many years ago and I expect Deep Fritz >>>also to get more than 2650 against humans. >>> >> >>I don't believe this for one second. Based on watching humans play programs >>like DF, DJ and Tiger in human events. There is just nothing to suggest that >>this is the case at all. >> >>And I don't see anything at all that would suggest that DF would be better than >>2655 today. But even if it is, what does the factor of 50 say about the DB2 >>evaluation? What does the factor of 200 say about DB2's speed? >> >>In one thread you said that Stobor at some ridiculous speed would whip all >>programs today. And you think Stobor is even a small fraction of what DB2 >>is all about? Knowledge? Search quality? Etc? >> >>You have to pick one side of the speed argument and stay there. > >Deeper blue had not the ridicilious speed that I suggested and I guess that the >search algorithm of it is better than deeper blue. You think the search algorithm of Stobor is better than deep blue 2? Based on exactly what? Stobor has nothing like singular extensions, for one example. It has no real chess knowledge compared to other programs or deep blue 2. I don't see the basis for the statement. Don't forget that DB2 _could_ peak at 1B nodes per second. That is a pretty ridiculous speed, IMHO. When the rest of use are happy to peak at 1M. > >I also doubt if the 200M nodes per seconds were used efficiently and it is >possible that 200M nodes of deeper blue is eqvivalent only to 20M nodes of Deep >thought(one possible problem is searching the same node by more than one >processor and I can imagine that there were more problems,for example changing >the search algorithm to something slower that is more simple to use). I don't know what that means. 200M on an SMP box is not as good as 200M on a single-cpu, I agree. However, DF and DJ are also running on SMP boxes, and they are _nowhere_ near 8x faster there either. Both have the same issues to handle. But DB's parallel search was developed for 10 years, while DF/DJ are much younger, and almost certainly less efficient. Hsu didn't change to a "simpler search" at all. This is covered in his thesis quite well. It is probably the most sophisticated parallel search ever attempted, with some roots in the Cray Blitz parallel search, but modified because of the 2-level parallelism he had to work with. >> >> >> >> >> >>>I also doubt the claim that Deeper blue was >>>100 times faster and smarter than deep thought. >> >>There is no "claim". Chiptest searched 350K nodes per second or so. Deep >>Thought used from 2 to 14 (14 hardly ever worked) of those chips. Hsu claims >>that DT was doing about 2M in 1992. He also has clearly stated that DB2 was >>doing 200M. That is 100X faster. > >this is the point that I am not sure that 200M is 100x faster I believe he factored in some "loss" in his 200M number. Otherwise we should see a far higher NPS. He reported that the SP2 could drive the 480 chess chips at 70% duty cycle in typical positions. that factors out to 700M nodes per second. What drops it to 200M if the chess processors are running at 700M all the time, according to his reports? All I can guess is that he factors in search overhead to say "My DB2 nps figure would compare favorably to a single cpu machine searching at 200M nodes per second." That is the only way I can see to arrive at 200M from 700M. If it were Crafty, that 200M would be much higher if the total was 700M. But he did have some issues to work around that cost him some speed. However, I take the 200M at face value, based on some simple analysis of what the machine could do. IE if he wanted to compare the thing to Crafty, I say my speed is roughly 1M nodes per second. Using that approach, he would say DB2 searched at 1B nodes per second. Because both are peak values, not factoring overhead caused by searching duplicate nodes on different processors, which _everybody_ has to do. >> >>He also stated that he went to a 25X as dense fabrication process, but stayed >>with the same size chip die. He then said that for DB2 the evaluation had >>increased to about 2/3 of the total chip area. That equates to at _least_ >>25 times more evaluation and that only accounts for the denser circuitry. >> >>So what is there to doubt? >> >> >> >> >> >>> >>>My impression based on the some analysis of the logfiles is different. >>>I did not give programs hours to analyze every position >>>that Deeper blue pondered and I may change my opinion >>>if a comparison between the logfiles and analysis of top programs >>>can show me that deeper blue is really at least 100 times faster >>>than part of the top programs in significant number of positions. >> >>How can you tell? DB2 didn't even _count_ nodes it searched. It would >>take as long to count 'em as it would to search 'em (A direct quote that >>dates back to Ken Thompson). You can't see the last 1/3 (or more) of its >>PV. It is hard to compare knowledge unless you find positions that have a >>correct positional move that must be played and see whether everyone else >>can find the correct move or not. > >I believe that if we analyze the games carefully we are going to find some non >trivial positinal move that must be played. > I hope you do... that will be interesting. Although I don't know how we can compare that to any other program. If DB finds something positionally at depth=N, another program can find it positionally at depth=X where X is larger or smaller than N, depending on the weights... >If we search the hundreds of positions and everyone for hours then I >expect to find at least 10 positions when it happens. > >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.