Computer Chess Club Archives


Search

Terms

Messages

Subject: Re: Hsu Presents a Paper at

Author: Roberto Waldteufel

Date: 15:35:05 06/25/98

Go up one level in this thread



On June 25, 1998 at 14:02:02, Robert Hyatt wrote:

>On June 25, 1998 at 13:56:05, Robert Hyatt wrote:
>
>>On June 25, 1998 at 13:46:06, Roberto Waldteufel wrote:
>>
>>>
>>>On June 25, 1998 at 12:03:11, Robert Hyatt wrote:
>>>
>>>>On June 25, 1998 at 04:54:02, Roberto Waldteufel wrote:
>>>>
>>>>>
>>>>>Here's another question about Hsu's chess chip. I seem to recall reading some
>>>>>time ago that Hsu was considering a commercial release of his chip. Does anyone
>>>>>know anything more about this? If the chip were to become available, how could I
>>>>>use it in conjunction with a PC? would the fixed depth not be "out of sync" for
>>>>>the speed of, eg a Pentium 333Mhz if it was designed to work with a
>>>>>supercomputer, or can the fixed depth be adjusted to redo the balancing act in
>>>>>the new environment? If it were possible, I would be very interested in
>>>>>experimenting with this sort of hardware coupling. I assume that it would extend
>>>>>the depth to which a program could search by something like 4 extra plies within
>>>>>the same time. This would surely improve the strength of the PC ches programs
>>>>>quite a lot!
>>>>>
>>>>>Roberto
>>>>
>>>>there really isn't a "fixed depth".  This is a parameter you set in the chess
>>>>processor when you fire it up.  You simply tell it to search to a depth of your
>>>>choosing, although this depth has to be closely matched to the speed of the host
>>>>machine as I mentioned in the past.  Since the chess processor has a fixed clock
>>>>speed of 2.4M nodes per second, you have to choose how deep it searches to
>>>>balance that speed against the host machine that has to provide it positions to
>>>>search.  Set the depth too high and the host will waste time waiting on the
>>>>chess processor.  Set the depth too low, and the host will have to search an
>>>>extra ply deeper to be able to provide the chess hardware positions quickly
>>>>enough to keep it busy.
>>>>
>>>>If you study this, what you find is an interesting phenomonon.  If you tell the
>>>>chess processor to search 1 ply deeper, then the host processor has to search
>>>>one ply shallower, or it will provide positions faster than the chess processor
>>>>can search them and return the results.  If you tell the chess processor to
>>>>search one ply shallower, then the host processor has to search one ply deeper
>>>>to provide it positions to search fast enough.  In short, for a given host,
>>>>there is a very definite max depth you can search.  You can do more in the
>>>>hardware, or more in the software.  Software is generally better as you can
>>>>change the search, tweak with the extensions, without changing the chess
>>>>processors at all...
>>>>
>>>>Hope that explains what they are doing...
>>>
>>>Thanks for the info. I wonder if the theoretical limit problem could be overcome
>>>in the same way that they did with Deeper Blue, ie many chess processors in
>>>parallel, or would this not be practical for a pentium host?
>>>
>>>Best wishes,
>>>
>>>Roberto
>>
>>if life were only so simple.  :)
>>
>>the point is, for any specific host machine, there is an optimal number of
>>chess processors for a host machine.  If you get too many chess processors,
>>the host can't keep them busy unless you increase the depth on each chess
>>processor by a ply so that the search on the hardware takes longer.  But if
>>you do this, the host can't keep up unless you reduce the depth by 1 ply, so
>>you gain *nothing*.
>>
>>There is a constant ratio between the host processor and the number of chess
>>processors.  To go faster, you can double both, or to get cheaper you can divide
>>both by 2, but you can't vary the number of chess processors without varying the
>>number of host processors, *or* (perhaps) getting a single host processor that
>>is twice as fast, which would let you keep 2x the chess processors busy.
>>
>>that's why they used a Scalable SP machine...  if they want to double the number
>>of chess processors (to double their NPS) they have to double the host as well.
>>
>>Bob
>
>
>just thought of another way to explain this:
>
>think of the complete tree laid on it's side as a graph, with the root position
>on the left, and tip positions to the right.  You have two pieces of hardware,
>the host and the chess processor.  The chess processor is going to search the
>positions on the right side of this graph, and the host is going to search the
>positions on the left side.  Draw a line vertically to separate the tree into
>the left-half and right-half.  Where do you draw this line so that what is to
>the left is searched on the host, and what is on the right is searched on the
>chess procssor?
>
>if you draw it too far to the right, the host does more work and gives the chess
>processors tiny little trees to search.  If you move it to the left, the host
>searches a pretty small tree, but the work done by the chess processor is going
>to be proportionally larger, because all that stuff has to be searched to reach
>the same depth.
>
>So, in such a case, the host is waiting on the chess processor, while in the
>first case, chess processors are waiting on the host.  Ideal condition is to
>have *both* host *and* chess processor busy 100% of the time.  That occurs when
>you get that vertical line in just the right place, so that the host can just
>barely keep up with supplying the chess processor with positions, and the chess
>processor is just barely keeping up with the positions the host is sending it.
>
>There you get maximal performance.  Move the line left or right, and one will
>start waiting on the other.  With a loss of performance...
>
>Maybe that makes more sense that way?
>
>Bob

Hi Bob,

Thanks, I understand fine now. Even so, if the ratio is fairly close to the
optimum, the overall effect compared to host alone would surely be a huge
speedup, don't you think? I'm just hungry (even greedy!) for the most speed I
can get. I think that no matter how deep I manage to search, I will always have
an eye open for some way achieve that elusive extra ply. Perhaps the Alpha chip
is a more practical (ie available) way to upgrade my hardware from Pentium, but
I would imagine a chess processor (or several) would be the most dramatic of
improvements.

Best Wishes,

Roberto



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.