Computer Chess Club Archives


Search

Terms

Messages

Subject: Re: misunderstanding about PC cards (deep blue)

Author: Robert Hyatt

Date: 10:59:55 06/28/98

Go up one level in this thread


On June 28, 1998 at 13:45:09, Tom Kerrigan wrote:

>On June 28, 1998 at 06:09:47, Vincent Diepeveen wrote:
>
>>So suppose IBM permits certain people to try to put cpu at card in PC,
>>what is your conclusion about it, is it very easy possible considering
>>communication speed?
>
>I think communication speed is a total non-issue.
>I just read one of Hyatt's posts saying a chip searches 2.5M NPS.
>If a search takes 100,000 nodes, that's one search in 0.04 seconds.
>Now, consider a chess program on the PC that searches, say, 300k NPS.
>The same search would take it 0.33 seconds.
>Thus, if a program uses a DB chip, it's possible (even likely) that it would go
>8 times faster than it would without. Even if the communication overhead doubles
>the amount of time it takes the DB chip to complete a search (impossible in my
>mind), it would still make the program 4 times faster. And that's with the
>chip's fancy evaluation, mind you.
>
>Cheers,
>Tom


the 2.4M came directly from Hsu, although he did say that they used some that
were slower (older or less reliable at 24mhz I don't know).  But the "best" of
the chips run at 24mhz, with 10 cycles / node, total, although you can't try to
say "8 for eval, 3 for move generation, 5 for alpha/beta, etc, because lots of
that is pipelined in parallel).

However, I know how I'd use a gadget that can search 2.4M nodes per second (or
even better a PCI card with 8 of those just like the real DB boards) where I
could search 20M nodes per sec.  I would divide the tree at the point where
1/200th of it is searched by the PC (at 100K nodes per second) which could just
barely feed the chess processors that can search 200X faster.  And it would
definitely be impressive, and communication would be trivial, so long as you
don't try to search a single node per update to the chess processor which would
be silly of course.

So I agree... communication is a non-issue if it is on the PCI or VME or EISA
or whatever bus.  If it plugs into a serial port, that would be different
probably...



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.