Computer Chess Club Archives


Search

Terms

Messages

Subject: Deep Blue chip talk

Author: David Fotland

Date: 07:46:59 08/19/98



I attended Feng-hsiung Hsu's talk on Deep Blue at Hot Chips 10, and
asked a few questions afterwards.  He described the chip and its
performance.

Here's what he said:

Search speed is important.  He thinks that "Normalized search depth"
needs to be 13 ply to beat Kasparov.  He put Hitech at 9 ply, Deep
Thought at 10 ply, Deep Blue prototype at 11 ply.

But speed alone is not enough to beat GM in a match, since GM will
prepare and learn program's weaknesses.  Need to have very few
weaknesses, and only weaknesses that are difficult exploit.

Chip speed is secondary.  Integration level and chess knowledge is
paramount.  Encapsulate every chess evaluation term from chess books.
Create terms to deal with every known computer weakness.  Put everything
on one chip.

Deep blue had 2 SP2 Frames.  Each has 15 SP2 nodes (30 total).
Each node two deep blue boards.  Each board has 8 deep blue chips.
200 million chess positions per second average.  he estimated that
each chess position was equivalent to about 10,000 instructions
on a general purpose CPU.

the chip contains:
- move generator
- move stack
- repetition detector
- evaluation function (fast and slow)
- alpha-beta search control
- connected with 19 bit wide move bus

Evaluation is about 2/3 of the chip area.  Move generation is about
3/4 of the rest.

Move generator is similar to Deep Thought.  8x8 combinational array.
generates a move in 2 clocks.  Can generate checking and check evasion
moves (used in quiescent search).

The fast evaluation is one cycle and includes all the big terms.
The slow evaluation takes 8 to 11 cycles, iterating across the columns
of the chessboard, then combining results.  Many small rams are
used in the evaluation.

The fast evaluation includes material, engame tables and logic for
draw detection, endgame king and pawn race, and piece placement tables.
The piece placement tables take the from square, to square, attacking
piece type, victim piece type, and do table lookup to get a value.
The tables are small.  1024x10 bits indexed by attacking piece and
to square, 512x10 bits indexed by attacking piece and victim piece
for each color.

The endgame tables are only KP vs K, KR vs KP, KQ vs KP, and
KRP vs KP.

The slow evaluation scans one column per clock and accumulates
results.  It has special logic to figure out development and
pins.  It has many small rams for different evaluation terms
(exact number not specified).

The chip is 0.6 microm CMOS (old technology).
1.5 million transistors, 1 watt, 25 Mhz clock.
2.5 million chess positions per second per chip sustained (10 clocks ave).
Single chip plays at strong grandmaster level.

480 chip system played against Kasparov, and sustained about
200 million chess positions per second (less than 1/5 full speed).

He estimated the publicity was worth $2 billion to IBM.

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.

He said the key to success is the evaluation function.  He tested
against some micro programs using a single chip, slowed down by
5-10X, and beat commercial programs 10-0.  He said this was because
of the work they did to avoid typical computer chess weaknesses.

The opening book has 3 parts:
- grandmaster book designed to get to interesting positions
- autogenerated book from game collection
- in positions where the autogenerated book has low counts and
poor statistics, a random bonus is generated to vary play.

Hsu is negotiating with IBM for rights to the chip, and plans to
commercialize it.

David Fotland



This page took 0.04 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.