Author: Robert Hyatt
Date: 13:07:36 12/07/99
Go up one level in this thread
On December 07, 1999 at 15:25:11, Vincent Diepeveen wrote: >On December 07, 1999 at 14:31:17, Robert Hyatt wrote: > >>On December 07, 1999 at 09:02:37, Vincent Diepeveen wrote: >> >>>On December 06, 1999 at 15:33:03, Robert Hyatt wrote: >>> >>>>On December 06, 1999 at 13:00:56, Georg v. Zimmermann wrote: >>>> >>>>>>A thousand fold increase would be >>>>>>what, an additional 6 ply search in the same time? >>>>> >>>>>Lets do some math. 40^x = 1000, 40log 1000 = x, x = 10log1000 / 10log40, x = >>>>>3/10log40 = 3 / 1.5 = 1.9 >>>>> >>>>>I think it gets you "1.9 ply" deeper if you do brute force. Now we need someone >>>>>to tell us how much that is if you add HT and other modern wunder drugs. >>>>>But I would be very very suprised if you'd reach +6ply. >>>> >>>> >>>>DB has an effective branching factor of roughly 6, about the same as Cray >>>>Blitz, which didn't use R=2/recursive null move. Log6(1000) is at most 4, >>>>so it would get about 4 plies deeper. Certainly nothing to sneeze at... >>> >>>see different post of me. DB may be happy with a b.f. from 10.33 >>> >>>>But then again, this math is really wrong, because for each cpu, DB used >>>>16 chess processors. Each chess processor could search about 2.4M nodes per >>>>second (they used almost 500 for DB2 the last match). With one million >>>>processors, they would then have 16M chess processors, and would be >>>>searching about 40,000,000,000,000 nodes per second. At about 1 billion >>>>(max) for DB2, this would be 40,000 times faster. and log6(40000) is 6, >>>>so they could hit about 6 plies deeper. Very dangerous box... >>> >>>the more processors the smaller the speedup. just attaching all processors >>>to the search might take a few minutes. >>> >>>Note that HSU writes that they got very close to 1 billion positions a >>>second but never hit the magic 1 billion positions a second number. >>> >>>Vincent >> >> >>Sure.... hitting 1B is not easy when you have _just enough_ chess processors >>to peak at 1B. But to hit 1B requires perfect speed-matching between the >>chess processors and the SP, which doesn't happen. I think he said that the >>chess processors were running at about 70% of max speed because of this. And >>he also claims 30% efficiency (in a linear way) in his parallel search. Which >>means that no matter how many processors he adds, he gets about 30% of each one. >> >>As far as branching factor, he uses normal alpha/beta, so I have no idea where >>you would get 10+. > >See a post some higher. > >axb5 was a fail low. way over 3 minutes. > >800M * 180 seconds = 144 * 10^9 nodes. >11th root out of that is 10.33 > >simple nah? > >but the reason why is obvious: > - normal alpha beta without good move ordering is a crime > - no hashtables > - in the normal search DB did a lot of extensions > blowing up the search. extensions especially blow up the > search if you don't nullmove. > - i don't believe his 30% claim unless he was minimaxing. > >Vincent 1. your math doesn't work.. because you have _no_ idea how many nodes it takes him to search a 10 ply tree. Effective branching factor = 11 ply time / 10 ply time. Anything else is a pure guess. I see nothing that they do that would drive the EBF beyond sqrt(38) which is roughly what alpha/beta is supposed to be. 2. Move ordering that they do is very similar to ours... Particularly in the software (first 8 plies + all the extensions). Move ordering in the hardware is more simplistic of course, using MVV/LVA to sort captures. 3. They have hashing in software... but not in hardware. The hardware supports hashing, but he lacked time to design/build a big multi-port memory for each group of 16 cpus... 4. I believe anything he says until I see evidence that he is misleading everyone. So far it hasn't happened. They did a lot of testing and the 30% seemed pretty accurate. Not good, of course... but 30% of 512 is still a huge speed-up...
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.