Author: Walter Faxon
Date: 13:14:50 04/11/05
Go up one level in this thread
On April 09, 2005 at 07:49:03, Tony Petters wrote: >On April 08, 2005 at 21:06:02, Walter Faxon wrote: > >>On April 08, 2005 at 18:16:15, Tony Petters wrote: >> >>>On April 08, 2005 at 15:50:45, Walter Faxon wrote: >>> >>>>Just looking around and found this >>>>(http://sccs.muldermedia.de/index.php?section=links&subsection=deepblue) record >>>>of a talk by Feng-hsiung Hsu at Hot Chips 10 (1998): >>>> >>>>"He thinks state of the art process can give 30 million chess positions per >>>>second in a single chip today. A small array of such chips plugged into a PC >>>>could beat Kasparov. In a few years (0.18u) a single chip could be as fast as >>>>the entire Deep Blue machine." >>>> >>>>Today MOSIS can supply custom 0.18u chips >>>>(http://www.mosis.org/Orders/Prices/price-list-domestic.html#ami12). I estimate >>>>that their undiscounted price for a single (large) 150mm^2 chip to be about >>>>$236,625. >>>> >>>>Hmmm... I think I'll wait and buy it used! >>>> >>>>-- Walter >>> >>>Are you saying that a single CPU can be purchased that makes 30 millions calcs >>>per second is available for $ 236,000 ? >>> >>>DB was clocked at 200 million calcs per second, so why would it be able to beat >>>Kasparov ? >>> >>>Cordially >> >> >>Not 30 million standard computer instructions, rather 30 million chess positions >>per second. That's move generation, position updating, and a very complex >>static evaluator, plus search control. Equivalent to a thousand or probably >>several thousand simple integer instructions per position. That was possible to >>design into a single chip in 1998. A better chip could be designed and built >>today. You cannot now buy either off the shelf. >> >>M NPS = Million Nodes (positions) Per Second: >> >>1997 -- Deeper Blue, older technology -- 1M NPS/chip x 200 chips = 200M NPS >>1998 -- then-current chip technology -- 30M NPS/chip x 7 chips = 210M NPS >>2005 -- 0.18u chip technology -- 200M NPS/chip = 200M NPS on a single chip >> >>Dr. Hsu assumes 200M NPS is enough to beat a Kasparov, since it was in 1997. I >>agree since the programming, which has been criticized, can only be improved. >>Emerson tan is correct that IBM owns all the DB software, but with enough money >>it can be bought or duplicated. (Probably not bought if by an IBM competitor.) >> >>However, Dr. Hyatt has written elsewhere that due to the increasing power of >>standard PCs and programs, there isn't enough of a market to make selling these >>chess chips, however packaged, a paying proposition. It would be a publicity >>stunt, like DB itself. >> >>-- Walter > > >That is interesting Walter, is this chip being used by a computer chess program >anywhere on the Internet ? > >2005 -- 0.18u chip technology -- 200M NPS/chip = 200M NPS on a single chip > >Do u have the exact URL link to this chip, I could not find it at your orginal >link posted. > >Yes, this would be nice to have in a PC, it would certainly increase the power >of chess programs and computer games in general !! No URL. See my quote above: "You cannot now buy either off the shelf." The original post would allow one to roughly estimate what it might cost to produce a Deep Blue-like machine today. That's a 1/4 million to produce a single chip, assuming you had a design for it. Maybe a like amount of money to design it -- unless you want to spend a few semesters learning integrated circuit design and design it yourself, but you would need access to commercial IC design tools. A newly-designed chip probably would have problems and you'd have to go through the design-produce-test cycle a couple of times (at 1/4 million a pop). Then there's the matter of programming it. At any rate, a lot of work plus a lot of money. But, importantly: A lot _less_ money than IBM used. -- Walter
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.