Author: Robert Hyatt
Date: 13:46:54 11/09/98
Go up one level in this thread
On November 09, 1998 at 13:46:58, Ed Schröder wrote: >>I *know* I don't write that confusing. What I said was this: DT had lots >>of *known* design weaknesses in the chess-specific hardware, from the evaluation >>to the search itself, and yet it rolled over every program around, from the >>micros right on up to the supercomputers, *in spite of* those substantial >>weaknesses. The hardware was redesigned at least twice in a major way, from >>the late "deep thought" hardware to Deep Blue I, to the chip used in the last >>match (I'll refer to it as DB II). > >>I brought this up because *you* were mixing micros of today with deep thought >>of 7-8 years ago. From experience, Cray Blitz of today still outplays any >>micro I know of, > >Can you post the CB games then? > >Or at least give some results? I didn't save the game... I published them in r.g.c about 4 years ago... about 50 games vs genius at various handicaps... IE at the time, the machine I was using was a C90 with 16 processors. My handicapping had cray blitz using one cpu, one second per move... and I tried genius at 5 secs/move... that was an 80X handicap when you factor 5:1 plus the 15 missing cpus. This was a total blowout. CB didn't lose a single game and most were simple tactical blowouts that lasted less than 30 moves. (Note: genius on a 486/66 and then 486/100 later). I ran genius up to 30 seconds per move and it started scoring about one of every 10-15 games. When I gave it a minute, still with cb at 1 cpu and 1 processor, the match became more interesting... CB was still winnint 3 of every 4 or 5 games... but I didn't play many at this time control as it was slow and cray time was not easy to get. And yes, that was old hardware... CB today is about 6X faster than CB of 4 years ago, because of the T90 with 32 processors at 2x the speed, plus some more enhancements that CB seems to like... My assumptions, then, are based on CB maintaining about the same hardware speed advantage today that it had 4 years ago, when you assume a doubling every 1.5 years... And no, I can't say a lot about *long* time control games... because I couldn't (and still can't) play standard type time controls using CB. I played as I did back then because I can run up to 5 minutes of CPU time, then I get kicked off and have to relog back in and start over or resume the game in the middle. So I chose to keep CB's time limit such that it would almost never need 5 total cpu minutes... > > >>yet I know exactly how it did against Deep Thought and >>"deep blue prototype". > >>But more importantly, I also know just how much better the current Deep Blue II >>is when compared to those older versions of their hardware. And their old >>hardware was so overwhelming... If you actually think a micro of today could >>beat Deep Thought, there's not much I can say to change your mind. And that's >>ok. But the difference between you and I is that *I* have sat across the table >>from them on several occasions and seen what they can do. And I have sat across >>the table from micro-based programs many times to know how *they* played as >>well. And the gap between the two looks somewhat like the Grand Canyon. And >>then there is deep blue, which is far stronger (faster, smarter, etc.) than the >>program I had to play, and *no* I don't think it is weaker today than in 1994 >>when they forfeited round 1 at the ACM event and *still* won the event. > >>My summary: I don't think a micro of today could beat the deep thought of the >>early 1990's except for the random statisitical oddity. A match? *zero* >>chance. A single game? A vanishingly small chance, but a chance (Fritz proved >>this at Hong Kong of course). > > >>>Very revealing. I was wondering when this secret basement match would be >>>mentioned again. In a sense it's appropriate that DB's reputation would be based >>>on such rumours. I remember it was discussed "quite a bit" by Hsu and Campbell: >>>they were trying to remember who were the opponents, what was the hardware :) > > >>There wasn't any doubt about the opponents... Rebel and Genius. There was >>some discussion about the "version". And even the actual CPU speed. But the >>hardware wasn't "cheesy" while they were slowed down to 1/500th of their normal >>speed. So there's not much room for complaint there, IMHO... > >I remember the following: they used Rebel7 and Genius 3 or 4. Pc a P90. Time >control 30 seconds average. No games. Only the 10-0. I seem to remember that all games weren't 30 seconds... but I don't remember the specifics any longer... 30 secs might be a good number for the average of the 10 games perhaps... > > >>I trust their results and statements. And will always do so until I see them >>say something that I *know* to be false... I doubt that will happen however... > > >>Two years ago I had a neighbor that owned a classic Ferrari 308. I *knew* it >>would blow the doors off of my truck without ever having to go to the race track >>with him to run the 1/4 mile... There are some things that you can "know" even >>without a competition... > >Not quite so. Until they play I would say. The last time they played turned out >a disaster for them. Also then they had this enormous speed advantage (your >Ferrari 308 so to say). > wrong games.. :) The *last* time they played they astounded the world as I recall. :) >I remember our RGCC discussion that (I believe it was the Fritz game) we >caught the program on a horrible horizon effect of just 10 plies. Remember? > >- Ed - it happens.... but remember the position I posted here about a year or go or so from the Orlando ACM event... where they found this incredibly deep way to trap a bishop... and they found it 10 *moves* before Cray Blitz began to get an inkling something was wrong... And remember *both* of these cases were from a machine doing about 2M nodes per second... a couple of hundred times slower than the DB that played Kasparov...
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.