Author: Robert Hyatt
Date: 19:26:45 09/10/01
Go up one level in this thread
On September 10, 2001 at 17:29:06, Uri Blass wrote: >On September 10, 2001 at 16:34:37, Robert Hyatt wrote: > >>On September 10, 2001 at 16:06:40, Uri Blass wrote: >> >>>On September 10, 2001 at 15:44:05, Robert Hyatt wrote: >>> >>>>On September 10, 2001 at 15:08:38, Uri Blass wrote: >>>> >>>>> >>>>>the game was Deep thought's game and not Deeper blue's game so it was not >>>>>200Xfaster than yours >>>> >>>>At that event, we were probably running on a Cray XMP I would guess. I will >>>>try to look at my old tournament booklets to see exactly what we used. If >>>>it was an XMP, which is likely, then we were doing maybe 80K nodes per >>>>second if we were lucky. >>> >>>I thought that Cray blitz could search 7M nodes per second. >> >> >>This was on 1995 hardware (the T932). The game vs deep thought was well prior >>to that hardware if I recall correctly. I am trying to dig thru a really thick >>file to see if I can find out what we were using for that event. But it >>definitely was not a T90 as we never played on a T90 in any competition. The >>best hardware we used was a C90 which could hit about 500K nodes per second >>peak. >> >> >> >> >>>If it is not the case then I do not understand the reason that you believe that >>>cray blitz (7M nodes per second) was weaker than Deep thought. >> >>I don't compare 7M cray blitz to DT. the 7M CB was in the same time-frame >>as the DB/DB2 machines. And should be compared to them. > >I remember that one of your claims in order to convince people that Deep thought >was strong was the fact that it defeated Cray blitz when Cray blitz is better >than Crafty based on your games. > >If the real Cray blitz with 7M per second was never used in tournaments then >the fact that Deep thought beated Cray blitz is not relevant > >> >> >> Ok.. Please pay careful attention for a few minutes. DT beat Cray Blitz on the best cray hardware available at the time. The last time we played them we were running on a C90 at something around 1-2M nodes per second. The statement that you and many others have made is "programs of the 1980's and early 1990's are nowhere near today's programs, regardless of how fast they go. I simply ran Cray Blitz on a current Cray, which happens to be maybe 3x faster than the last machine DT beat us on. If you think a factor of 3 is huge to a program with a branching factor of 5+, then you are mistaken. And if you think that there is no way to draw conclusions based on this match, you are mistaken again. If Cray Blitz was just a "fast/dumb program" then that extra speed would make little difference, in theory. Deep Thought was very strong. Because Cray Blitz was also very strong. Against both humans and computers. It registered the first win vs a chess master on record. It registered the first win of an "open section" tournament on record. It also won a couple of WCCC events along the way. It's credentials are unimpeachable. That deep thought beat it at every turn says something about them. That they beat it with no real difficulty, (the first chiptest game really doesn't count here) says even more. >> >> >>> >>>Was there a single game between the Cray with no bug and 7M nodes per second and >>>Deep thought? >> >>Nope. No games with the T90 against anyone other than the blitz games vs >>Crafty. >> >> >> >>> >>><snipped> >>>>So what? I would _much_ prefer to have hardware designed specifically to >>>>tackle the chess problem, as opposed to a general-purpose computer. There >>>>is really no comparison between the two... >>> >>>They did not have the hardware. >>>The hardware was ready only short time before the match so they could not >>>do tests to optimize their code. >> >>Sure they could. They had deep thought hardware while waiting on the faster >>DB chips. They had DB1 chips while waiting on the faster DB2 chips... > >I thought that part of the things were hardware implementation and they could >not do it on their older hardware for similiar reasons that you cannot do the >same things on Crafty and Cray blitz. Some things they couldn't do (new eval terms could not be tested on old hardware, for instance). But the search was all software, and it received a lot of work as well. Remember, they were doing this stuff from 1987, for the next 10 years... Crafty vs Cray Blitz is a different matter. I can do _anything_ in crafty that I did in Cray Blitz. It is just too expensive to be practical, in some cases. But for testing, I could certainly put up with the slower performance, knowing I would get that back on the right hardware. >> >> >> >> >>> >>>I also understod that the code could not work on slow hardware because the slow >>>hardware was not designed for it so they even could not test it on slow >>>hardware. >> >> >>See above... >> >>That was one reason Deep Thought was called "deep blue prototype" for a couple >>of ACM events. It was old (and slow) hardware, but running the new software >>stuff. >> >>> >>> >>>>> >>>>>Using available hardware that is 200 times faster and developing hardware are >>>>>different things. >>>> >>>>That doesn't compute to someone that has designed hardware. Hsu spent a lot >>>>of time developing the hardware. The _rest_ of the team spent a lot of time >>>>developing the code. >>> >>>When you cannot test the code on the hardware because the hardware is not >>>available you have problems that you do not have when the hardware is available >>>so saying that they were 100 rimes faster is unfair. >> >>100x faster is just 100x faster. Has nothing to do with debugging, testing, >>designing, or anything else. > >I meant that it is not a fair comparison to take only the speed difference in >comparing between Hsu and people like Amir Ban,Fransh morsh and other >programmers of top programs. Why? When I consider them all pretty equal otherwise, so far as programming skill and chess skill goes. Actually the DB guys probably have a serious edge in the chess skill stuff. But Programming-wise, I consider them all equal. And if one "equal" has hardware 100x faster than another "equal" then that person is more equal than the other. >> >> >> >>> >>>You also had problems with Cray blitz because of the fact that you could not >>>have enough time to test it and it lost games because of bugs. >>> >>>Uri >> >> >>Different issue. We could hardly get time on a 60 million dollar computer. >>They had all their old hardware 100% of the time. > >Do you say that the old hardware really could do everything that the new >hardware could do when the only difference is speed? > >Uri It depends on what is being tested. One big thing they obviously did in DB2 was to add some form of futility pruning. That could be done in software to test it. The new hardware would run faster, but that would be the only real difference. For evaluation, they probably had things they could not test until the hardware was available, although I am sure they had the same sort of software eval emulation code they used for the automatic tuning of deep thought. So they could at least have a good starting point for eval weights when the new hardware rolled in.
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.