Computer Chess Club Archives


Search

Terms

Messages

Subject: Re: Q&A with Feng-Hsiung Hsu

Author: Vincent Diepeveen

Date: 11:25:00 10/13/02

Go up one level in this thread


On October 13, 2002 at 12:00:39, Andrew Williams wrote:

It was answerred another 2 times that it's 12 ply simply,
excluding qsearch+extensions. So the hardware is inside that
12 ply, but the hardware depth FROM that 12 ply can be UP to
6 ply. So it isn't always searching 6 ply in hardware. It's
a variable depth which they can do in hardware. That explains
probably extensions around mainlines with captures very well
and do not forget that the chips could do up to 1 billion nodes
a second in theory.

It searched on average over a period of 3 minutes only 126 MLN
nodes from that. that's only 10% effective usage or so each
chip. That means that there are *always* loads of chips idling.

So getting extra processors to the PV is a very clever thing to
do then. Of course you give the chips a small search depth then.
2 to 3 ply is what Brutus was using at world champs 2002. Bigger
search depths in hardware were too inefficient.

You cannot use killermoves in hardware and you cannot use hashtables
and Hsu didn't use nullmove either. With his last so many plies pruning
(whatever name you want to give it, razoring, futility pruning) that means
that when searching the Principal variation, you have to give many processors
small search depths, because the pruning around principal variation is
a lot less than for the other moves.

It would be interesting to hear in a next chat how many plies the forward
pruning was in the hardware part. Here just doing 1 or 2 ply didn't
reduce anything. Yet i remember some statement from a talk at M$ from
Hsu recently where (was it Tom kerrigan) asked Hsu about forward pruning
in the hardware chips and he answerred he was doing that.

So what type of pruning he did there is not so interesting. Interesting is
that it saved them up to 90% nodes in hardware. That's very clever, because
chrilly concluded that without forward pruning in hardware, your tree gets
just TOO huge (because move ordering is near random).

Chrilly uses nullmove. If Deep Blue used something else there. Only
interesting thing from my viewpoint therefore is to know how many plies
they did this forward pruning in hardware.

My own experiments are not relevant here too much, because i use hashtable
and nullmove everywhere. So it is very well possible that without nullmove
the effect of forward pruning is way bigger than without.

>Several people asked:
>
>Question: What does "12(6)" mean in Deep Blue's logs?
>
>Dr Hsu said, "12(6) means 12 plies of brute force (not counting the search
>extensions & quiescence). 6 means the maximum hardware search depth allowed.
>this means that the PV could be up to 6 plies deeper before quiescence."
>
>Unfortunately questions were being fed through a third-party, so it wasn't
>possible to get a follow-up.
>
>Andrew



This page took 0.05 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.