Computer Chess Club Archives


Search

Terms

Messages

Subject: Re: DB chip?

Author: Robert Hyatt

Date: 06:20:07 05/13/99

Go up one level in this thread


On May 13, 1999 at 00:31:53, Mark Young wrote:

>On May 11, 1999 at 18:28:53, Robert Hyatt wrote:
>
>>On May 11, 1999 at 18:05:21, Charles L. Williams wrote:
>>
>>>On May 11, 1999 at 13:11:16, Robert Hyatt wrote:
>>>
>>>>On May 11, 1999 at 12:48:22, blass uri wrote:
>>>>
>>>>>
>>>>>On May 11, 1999 at 12:01:05, Robert Hyatt wrote:
>>>>>
>>>>>>On May 11, 1999 at 03:06:32, Dave Gomboc wrote:
>>>>>>
>>>>>>>On May 07, 1999 at 19:18:22, Robert Hyatt wrote:
>>>>>>>
>>>>>>>>On May 07, 1999 at 18:46:32, Dann Corbit wrote:
>>>>>>>>
>>>>>>>>>On May 07, 1999 at 17:48:48, vitor wrote:
>>>>>>>>>[snip]
>>>>>>>>>>this is off topic, but why didnt you ever try making a hardware version cray
>>>>>>>>>>blitz? or is that some future project? it seems cray blitz was always up against
>>>>>>>>>>hardware programs like belle ,hitech, deep thought.
>>>>>>>>>Of those machines, only deep thought had dedicated chess circuits.  The others
>>>>>>>>>were general purpose machines, running a computer program.  Just like Cray
>>>>>>>>>Blitz.  Cray Blitz was more than a match for all except Deep Thought, which had
>>>>>>>>>specialized hardware.
>>>>>>>>>
>>>>>>>>>Why didn't Dr. Hyatt write special hardware circuits?  That would be a pretty
>>>>>>>>>expensive hobby.
>>>>>>>>
>>>>>>>>
>>>>>>>>actually they were _all_ hardware machines.  Belle was the first special-
>>>>>>>>purpose chess machine...  Hitech was next, built as a vlsi project at CMU,
>>>>>>>>and finally deep thought which also originated at CMU.  Cray Blitz was the
>>>>>>>>only general-purpose computer program of the group, although CB was highly
>>>>>>>>coupled to the Cray architecture, with a vectorized move generator, and a
>>>>>>>>very good parallel search...
>>>>>>>>
>>>>>>>>And you are right, in that except for deep thought, Cray Blitz was stronger
>>>>>>>>than the others...
>>>>>>>
>>>>>>>I was under the impression that Hitech was equal or (perhaps) slightly better
>>>>>>>than Cray Blitz.  It lost on tiebreak at the '86 WCCC to your program, but won
>>>>>>>some of the North American tournaments in the '84 through '88 range, didn't it?
>>>>>>>
>>>>>>>Dave
>>>>>>
>>>>>>
>>>>>>Berliner wanted everyone to believe this.  And in 1985 it was even true as we
>>>>>>were searching 80K nodes per second to hitech's 120K or so.  But in 1986 and
>>>>>>later, we were better.  In 1989 we were 5X faster due to newer hardware...
>>>>>>
>>>>>>HiTech won the 1985 ACM event, we won the 1986 WCCC event (and beat HiTech in
>>>>>>the final round to win, in fact).  I don't remember them winning anything beyond
>>>>>>that because in 1987 this pesky thing known as "chiptest" and then "deep
>>>>>>thought" was unveiled...  :)
>>>>>>
>>>>>>IMHO, HiTech was never "better" than CB.  It may have been as good.  But the
>>>>>>only 'down' time for Cray Blitz was the 1985 event where a poor change by me
>>>>>>produced some ugly pawn positional play that killed it in two games in 1985,
>>>>>>and in the second round of the 1986 WCCC before I found and excised the 4
>>>>>>lines of code that were killing it.
>>>>>>
>>>>>>After 1987 there was never any doubt who was best from that point forward,
>>>>>>the author being Hsu...
>>>>>
>>>>>I know that there is a doubt  about it
>>>>>some people(not me) believe that deep thought is not better than Fritz3(P90).
>>>>>
>>>>>They could prove to the public after they lost to Fritz that they are better
>>>>>than Fritz by playing 20 games between them and Fritz and doing the games public
>>>>>but they did not do it.
>>>>>
>>>>>Uri
>>>>
>>>>
>>>>Everyone should read Hsu's paper in IEEE Micro.  He mentions the 10-game match
>>>>that causes such an uproar of denials, and goes on to give results over a total
>>>>of 40 games...  and it is pretty eye-opening....
>>>>
>>>>Not to mention the fact that he may be ending computer chess as we know it by
>>>>releasing a pc-compatible version of the DB chip.  And for those that want to
>>>>talk about commercial programmers using this hardware, forget the idea, because
>>>>the concept is _flawed_.  This is DB evaluation, and DB search.  All that can
>>>>be modified is the first N plies of the search.  So trying to graft this on to
>>>>some other 'engine' only produces a new flavor of deep blue, not a new flavor
>>>>of the base engine.  The evaluation and last few plies of search are the heart
>>>>and soul of a chess program.  And in this case, the heart and soul is pure
>>>>deep blue.
>>>>
>>>>Things are going to change in a serious way before long...
>>>
>>>
>>>So what's the plan?  Will there be a DB chip on a card we can plug into our PCs?
>>>  It seems like this will help the programmers, by giving them something
>>>extremely strong as a reference for developing and tweaking their programs.  On
>>>the other hand, a chip is hardware, and not so easy to tweak.  It seems like a
>>>DB chip is advantageous to us.
>>>
>>>
>>>Chuck
>>
>>
>>Apparently there will be a PCI card that can be plugged into a PC just like
>>any other PCI card now (network cards, SCSI cards, etc.)  This will include
>>one or more DB chips (probably not public at present.)  I'd expect that a
>>single card with a single DB chip would likely sell for 200 bucks (US) or so
>>based on comments by Hsu in the past (IE we had a long conversation about this
>>in Cape May at the last ACM event a few years ago.)
>>
>>It will take him some time to (a) fab the newer DB chip, design the PCI
>>interface, (b) modify the current DB software part to work with the new PCI
>>hardware and on a pc platform, (c) do whatever else is needed to provide a
>>commercial-quality product interface.
>>
>>DB's chess processor is static in regard to what it can evaluate and how the
>>search is done, it is dynamic in that evaluation weights can be modified
>>easily or disabled (set to 0).
>>
>>Hsu estimated 30 million nodes per second on a single chess processor.  This
>>using the same 'approach' as the current DB chip, only using a more modern fab
>>process.  That would be an absolute killer...  and using multiple copies of
>>such a chip, a PC could quite easily search way over 100M nodes per second and
>>be as strong as DB was in 1997.
>
>This is amazing, what was the cost of the deep blue machine in 1997? Why has
>this technology gotten so cheap in just a few short years? What can we expect in
>another 2 to 4 years after deepblue jr? (a billion NPS??? or more)
>
>Going from say a top of the line machine doing 1 million nodes a sec.(Quad PIII
>with the right chess program to take advantage of the Quad chip computer) to
>over 100M nodes per second. Are we likely to see this kind of performence jump
>again in our lifetimes for home use? Is Deepblue jr the beginning of the end for
>chess programming?
>


Hsu is re-fabbing the DB chip.  The 1997 version searches about 2M nodes per
second on a single chip, while the new version will search about 36M nodes
per second (again single chip speed.)  That can be made very economically in
a 1,2,3 or 4 processor version.  1997 DB used 480 of these chips, and a 16 cpu
IBM SP machine that was very expensive to drive the whole thing.

The 200 dollar version will likely search 30-40M nodes per second, although it
should be possible to make 'faster' versions (say 4 cpus on a single PCI card)
at a somewhat higher price.  No idea how pricing will really end up.

And yes, it will change chess programming forever...





>>
>>It will be very interesting...  although the concept of computer chess is going
>>to change, since nothing else will be within a couple of orders of magnitude
>>of the strength of that thing...
>>
>>Again, most of this was discussed in the current issue of IEEE Micro, which you
>>should be able to find (at least) at a local university library.



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.