Author: Robert Hyatt
Date: 10:46:52 08/29/01
Go up one level in this thread
On August 29, 2001 at 13:14:41, José Antônio Fabiano Mendes wrote: > Cray Blitz vs Belle [1981 North American Computer Chess Championship] > The (in)famous position Dr. Hyatt referred to the other day: > [D]5r1k/6p1/1n2Q2p/4p3/8/7P/PP4PK/R1B1q3 w > And here Cray Blitz played 28.Qxb6 > Source: http://wondersmith.com/rants/howfar.htm That's the one. However, I should probably give this some context as even in 1981, Cray Blitz could easily spot this... It wouldn't find the draw too quickly, but a reasonable search would spot that taking the knight was no good. What happened was that for a reason I don't remember now, we had to switch from the Cray-1 at CRI's headquarters, to a Cray-1 at the university of minnesota. I don't recall whether it was for the entire tournament or just part of a game now. But for this game we were on the machine at UM. The problem with that machine was that (a) it was non-dedicated. We ran and shared time with a few other users; (b) it had no "timesharing" so we had to submit everything as a "batch job". To get a move from the cray, I had to edit a file that had all the moves up to this point, add the move the opponent just made, and submit the batch job to the cray. It wouldn't run immediately, but had to wait until memory was available. After thinking, the batch job output would be sent back to the front-end machine we used and we had to keep checking to see if it had arrived. Once it did, I opened the file and discovered CB's move. All very sloppy if you think about it. Absolutely no pondering, of course, based on the way we were submitting a new job for each move. And to allow for the slowness of the non-dedicated cray, plus the batch job submission/queueing, etc, we had to use 30 seconds of cpu time as the limit. That is why Qxb6 was actually played. Cray Blitz in 1981 had just been ported to the cray. No vectorized code or anything. And we were searching at the blinding speed of 1K nodes per second. Over the next 3 years we got this to over 10K on the same machine. Cray Blitz also had a very relaxed "that was easy" code that said "if the first move is better than all the rest (and QxN seems to win a knight) and it doesn't fail low, then play it after 1/3 of the target time." It therefore used 10 seconds of cpu time at 1K nodes per second, and fell into a hole. This position takes about 2 seconds to find Bxh6 on my quad xeon using Crafty. Last time we tried it on a T90, it took a fraction of a second with Cray Blitz. But in 1981, it simply wasn't fast enough. Note that I don't exactly agree with the content of the web page you gave above when it compares today's machines to the Cray-1 of 1981. Cray Blitz of 1981 was not a representative program for the capabilities of that machine. Harry Nelson joined the team because Chess 4.9 was searching 2600 nodes per second on a CDC Cyber 176, a machine that was at least 2x slower than the Cray. He couldn't understand why and he helped for almost 15 years in making that program faster and faster while teaching me the tricks of the trade in Cray assembly (Harry was/is acknowledged as the best assembler person in the world for the Cray hardware, he wrote performance analysis tools, manuals, how-tos, and so forth). So, as Paul Harvey says, "now you know the _rest_ of the story." Remember that this 1981 version of Cray Blitz, at 1K nodes per second, was the same program/machine that won the 1981 Mississippi State Closed Chess Championship event, beating a real USCF master in the final round. It impressed Ken Thompson because we had eval code (yes, even back then) for bishop and wrong rook pawn endings and we failed to fall into a trap the human set, while Belle couldn't see it at all. So it wasn't weak, as this was the first tournament loss of a rated chess master to a computer, so far as we knew. The game was published in Chess Life along with a story about it and log analysis taken from the log file of CB. It still seems amazing what we were doing with 1K nodes per second when today we are 1000X faster. However, the program _did_ have a large evaluation with a lot of chess knowledge. That saved the bacon a lot of times.
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.