Citation and evidence

Deep Blue

35 min full readUpdated 39 references

This article's verification

Report a problem with this article

Edit article
More

Use this article

Raw MarkdownExplore connections

Improve this page

Revision historyDiscuss this article

Topics

Artificial Intelligence

Cite this article

Deep Blue was a chess-playing supercomputer built by IBM that, on May 11, 1997, completed a 3.5 to 2.5 victory over world chess champion Garry Kasparov in a six-game rematch in New York City. IBM describes it as the first computer system to defeat a reigning world champion under standard tournament time controls [2][18]. It chose its moves by search rather than by learning: a 30-node IBM RS/6000 SP computer carrying 480 custom chess chips, which searched roughly 100 million to 200 million positions per second depending on the position, using alpha-beta search and an evaluation function built from about 8,000 hand-designed features [4]. The match drew worldwide attention and was, in IBM's words, "widely seen as a symbolic test of whether supercomputers were catching up to human intelligence" [2].

Deep Blue relied on computing power and specialized hardware rather than machine learning or neural networks. Murray Campbell, one of its developers, has described the dominant AI of that era as "good old-fashioned AI, or symbolic AI, which was based less on machine learning," and said that machine learning methods for game-playing programs were then "fairly primitive and not able to help us much in building Deep Blue" [14]. That makes Deep Blue a prominent example of symbolic AI's search-and-knowledge-engineering approach, applied to a single, well-defined domain. As Campbell put it in 2017, "you could say chess is an enormously complex game, and that's why it took us, as a field, 50 years of development to finally beat the world champion" [14].

Where did Deep Blue come from?

Deep Blue's roots trace back to Carnegie Mellon University. In 1985, Feng-hsiung Hsu, then a doctoral student there, began developing a chess machine called ChipTest [2][6]. Working with fellow graduate students Murray Campbell and Thomas Anantharaman, the project produced a stronger successor, Deep Thought, named after the computer in Douglas Adams's The Hitchhiker's Guide to the Galaxy [6][14]. Both machines were built around a single-chip move generator; ChipTest searched about 500,000 positions per second and Deep Thought about 700,000 [4].

In 1988, Deep Thought became, in the team's words, "the first chess machine to beat a Grandmaster in tournament play" [4]. The Computer History Museum names that grandmaster as Bent Larsen and notes that Carnegie Mellon's rival machine, Hitech, beat grandmaster Arnold Denker the same year [30]. Deep Thought also won the second Fredkin Intermediate Prize for grandmaster-level play, based on a rating above 2650 on the U.S. Chess Federation scale over 25 consecutive games [3]. In 1989, Kasparov beat Deep Thought in a two-game match [6].

In 1989-90, Anantharaman, Campbell and Hsu moved to IBM's Thomas J. Watson Research Center in Yorktown Heights, New York, to keep building a world-class chess machine [4]; IBM dates the hiring of Hsu and Campbell to 1989 [2]. The machine was renamed Deep Blue, which IBM calls "a nod to IBM's nickname, Big Blue" [2]. An intermediate system, Deep Thought 2, played public events from 1991 through 1995 and won the 1991 and 1994 ACM Computer Chess Championships [4].

The team grew beyond Hsu and Campbell. A. Joseph Hoane Jr. replaced Anantharaman in late 1990 [4] and became the principal software engineer, responsible for the parallel search [36]. IBM also names Jerry Brody and C. J. Tan among the contributors [2]; Tan was the team leader during the 1997 match [18]. Grandmaster Joel Benjamin beat the single-node Deep Blue Jr. 2-0 in one of three tournament-condition matches played before the 1996 Kasparov match, a result that "went a long way to convincing us that Joel Benjamin would be an excellent Grandmaster consultant," the team wrote [4]. The Computer History Museum says IBM hired him several months before the 1997 rematch [31].

What hardware did Deep Blue use?

Deep Blue's strength came from its custom hardware. The 1997 version that defeated Kasparov, which the team sometimes called Deep Blue II, was configured as follows:

ComponentSpecification
Base systemIBM RS/6000 SP, 30 nodes (30 processors), connected by a high-speed switch [4]
General processors28 nodes with 120 MHz P2SC processors and 2 nodes with 135 MHz P2SC processors; 1 GB of RAM and 4 GB of disk per node [4]
Operating systemAIX 4.2 [4]
Custom chess chips480 single-chip chess search engines, 16 per node, on two Micro Channel cards of eight chips each [3][4]
Chip fabrication0.6-micron CMOS; about 1.5 million transistors on a 1.4 cm x 1.4 cm die [3]
Speed per chip2 to 2.5 million positions per second [3][4]
System speedAbout 100 million positions per second in tactical positions and close to 200 million in quiet ones; 126 million on average in searches longer than one minute during the 1997 match; 330 million maximum sustained [4]
Theoretical peakAbout 1 billion positions per second, a "guaranteed not to exceed" figure [3]
IBM's published figures"32 processors" and 11.38 billion floating-point operations per second [2]
Physical formTwo towers [29]; NPR called it "the 1.4 ton machine" before the rematch [27]

Expanded article table

IBM's history page gives "32 processors," while the team's own technical paper describes a 30-node system [2][4]. The 1997 machine was about twice as fast as the 1996 version; Campbell said the team "more or less doubled the speed of the system" [14]. Deep Blue I ran on a 36-node SP with 216 chess chips, each searching about 1.6 to 2 million positions per second, for 50 to 100 million positions per second overall [4]. Campbell later said the 1997 Deep Blue "searched between 100 million and 200 million positions per second, depending on the type of position" [14].

How did the chess chips work?

Hsu divided each chess chip into four parts [3]:

BlockFunctionDetails
Move generatorGenerated moves for the current positionAn 8 x 8 array of combinatorial logic, "effectively a silicon chessboard," that generated captures first and could also generate checks, check evasions and certain attacking moves [4]
Smart-move stackA regular move stack plus a repetition detectorThe repetition detector, new in Deep Blue II, tracked the last 32 plies and could spot a legal move that would repeat a position [4]
Evaluation functionScored positionsA one-cycle "fast evaluation" plus a "slow evaluation" that scanned the board one column at a time; about 8,000 recognizable features with programmable weights [4]
Search controlRan the on-chip searchA minimum-window alpha-beta search, with a 16-bit datapath and three state machines [3]

Expanded article table

The chips had no transposition table, and their search was simpler than the software search running on the host processors [4]. An optional interface for an external FPGA could have added a transposition table, more search control and extra evaluation terms, but "due to time constraints, this capability was never used in Deep Blue" [4].

How was the search parallelized?

Deep Blue was organized in three layers. One SP processor was the master and the others were workers, and every node, the master included, hosted 16 chess chips [3][4]. In Hsu's example of a 12-ply search, the master searched the first four plies in software, all 30 nodes then searched four more plies in software, and the chess chips finished the last four plies, including quiescence search. The software handled less than 1 percent of the positions searched but controlled about two thirds of the search depth [3].

The team counted "over 500 processors" available to the game-tree search [4]. They estimated the parallel search's efficiency at about 8% in tactical positions and about 12% in quiet ones, and wrote that the parallel search code was largely untouched between the 1996 and 1997 matches [4].

How did Deep Blue play chess?

Deep Blue's playing method was very different from how grandmasters describe their own play. Where human players lean on intuition and pattern recognition, Deep Blue combined a fast, highly selective search with a large hand-built evaluation function [4][14].

At its core, Deep Blue used alpha-beta search, a refinement of the minimax algorithm used in game-playing programs. Alpha-beta search explores a game tree of moves, replies and counter-replies while pruning branches that cannot affect the final decision. Deep Blue drew on established ideas from earlier programs, including NegaScout, a refinement of alpha-beta search [4], while each chip implemented a minimum-window alpha-beta search [3].

Campbell has said the system searched "between six and eight pairs of moves," and "to a maximum of 20 or even more pairs in some situations" [14]. The team's paper reports that a three-minute search reached a full-width depth of 12.2 plies on average [4], and Hsu wrote that during the 1997 match the software extended the search to about 40 plies along forcing lines, even though the non-extended search reached only about 12 plies [3].

The search combined standard techniques with some of the team's own:

TechniqueHow Deep Blue used it
Iterative deepeningSearched progressively deeper, a standard idea Deep Blue inherited from earlier programs [4]
Quiescence searchDone on the chips, which ran a "complex quiescence search" with captures, checks and some attacking moves [4]
Singular extensionsSupported by the chess chip, along with other search extensions [4]
Dual credit with delayed extensionsThe software's selective search, first built for Deep Thought 2, which extended forcing and forced pairs of moves once enough "credit" had built up [4]
Transposition tableUsed in the software search only; the chips had none [4]
Null-move searchUsed to detect threats, not to prune: the team wrote that research on null-move pruning "did not provide us enough evidence to warrant implementation in the hardware search" [4]

Expanded article table

Evaluation Function

The evaluation function was where human chess knowledge entered the system. According to the team, "The Deep Blue evaluation function is essentially a sum of feature values." The chip recognized roughly 8,000 different patterns, each with an assigned value, ranging from very simple ones, such as a particular piece on a particular square, to very complex ones [4]. The redesigned 1997 chip went from around 6,400 features to over 8,000, a number of them added in response to specific problems seen in the 1996 Kasparov games and in test games against Benjamin [4]. The paper counts 54 registers and 8,096 table entries, for 8,150 settable parameters [4]. Campbell credited the richer evaluation for much of the improvement: "Part of the improvement between '96 and '97 is we detected more patterns in a chess position and could put values on them and therefore evaluate chess positions more accurately" [14].

Some feature values were static, set at the start of each search. Others were dynamic, scaled during the search according to the material on the board; king safety, passed pawns and pawn-structure defects are examples [4]. An "evaluation function generator" running on the master node set these values at the root of each search [4]. The paper's tables of registers and evaluation tables include:

Feature groupExamples from Deep Blue's evaluation tables and registers
Piece placementThe largest single table (1,024 entries): basic piece values adjusted for the square each piece stands on [4]
PawnsPawn structure, passed pawns, multiple pawns, pawn storm, pawn shelter [4]
King safetyPawn shelter, plus king-safety terms inside other features such as rooks on files [4]
Mobility and blockageMobility tables, self block, opponent block, back block [4]
Rooks and bishopsRooks on files, rooks on the seventh rank, bishop pair, bishops of opposite colors [4]
Tactical patternsPinned, hung and x-rayed pieces; knight, rook and queen traps; trapped bishops [4]
OtherDevelopment, contempt, side to move [4]

Expanded article table

The weights were mostly set by hand, and the team did not use machine learning to build the function. They did build two automated aids. One used hill-climbing to find "noisy" features in the Deep Blue I evaluation. The other used comparison training to analyze pawn-shelter weights; it showed that the hand-tuned weights were systematically too low, and they were raised before the 1997 match [4]. The weights could also be changed from software. Hsu wrote of the 1997 match: "we didn't use the hardware escape hatch, but we did change the weights for the positional features between games" [3].

Opening Book

Deep Blue's opening book was created by hand, mainly by Joel Benjamin, with help from grandmasters Nick De Firmian, John Fedorowicz and Miguel Illescas. It held about 4,000 positions, each checked by Deep Blue in overnight runs [4]. The team acknowledged that this "may seem surprisingly small," explaining that numerous openings had minimal preparation because of their confidence in the extended book [4].

The extended book, introduced in Deep Thought 2 in 1991, summarized a database of 700,000 grandmaster games. It gave bonuses or penalties to moves that grandmasters had played, based on factors such as how often a move was played, the strength of the players, how recent the games were, their results and annotators' comments. The bonus, which could be as high as half a pawn, offset the alpha-beta search window for that move, nudging Deep Blue toward the consensus of opening theory without replacing its own search. When one move's bonus was much larger than the others, Deep Blue could play it without searching [4].

Endgame Databases

Deep Blue used endgame databases, precomputed tables that give the game-theoretic result for every position with a given set of pieces. They covered all positions with five or fewer pieces (counting pawns as pieces), along with selected six-piece positions that included a pair of blocked pawns. The main sources were Ken Thompson's CD-ROMs and additional databases supplied by Lewis Stiller [4].

Database setWhere it was stored
4-piece and important 5-piece databasesReplicated on the local disk of each of the 30 nodes [4]
All other databases, including 6-pieceDuplicated on two 20-GB RAID disk arrays shared over the SP switch [4]

Expanded article table

Positions were stored at one bit each, marking a position as lost or not lost. A database hit was scored with a game-theoretic value plus a tie-breaker from the evaluation function, so Deep Blue would prefer drawn positions over lost ones, and among drawn positions, the one it evaluated best [4]. The databases were also used offline to check the endgame patterns stored in each chip's ROM [4].

Ken Thompson of Bell Labs began applying retrograde analysis to chess endgames in the mid-1970s, finished his five-piece databases in the mid-1980s, and starting in 1991 made the full set publicly available on four CDs [12].

The endgame databases did not play a critical role against Kasparov. In the 1997 match only game 4 approached an endgame that would have required them, and the chips' ROMs already had enough knowledge to recognize the rook-and-pawn-versus-rook draws that could have arisen [4].

What happened in the 1996 Kasparov match?

The first match between Deep Blue and Garry Kasparov began on February 10, 1996, at the Pennsylvania Convention Center in Philadelphia, and ended on February 17 [6]. It was a six-game match, and the time control initially required 40 moves in two hours [4].

GameResultNotes
Game 1Deep Blue winsIBM calls it the first win by a computer against a reigning world champion under regular time controls [2]
Game 2Kasparov wins73 moves [7]
Game 3Draw
Game 4Draw
Game 5Kasparov wins
Game 6Kasparov wins
Final ScoreKasparov 4, Deep Blue 2[4][6]

Expanded article table

The match was tied 2-2 after four games [4]. Then, in Hsu's words, "Kasparov pinpointed Deep Blue's weaknesses and won the remaining two games easily" [3]. In the final game he used anti-computer tactics, keeping the focus on long-term planning while Deep Blue made little progress [7]. The 36-node Deep Blue I played only these six games under tournament conditions [4].

The team spent the next year rebuilding the machine. They designed a new chess chip with a redesigned evaluation function, added hardware repetition detection, raised each chip's speed to 2 to 2.5 million positions per second, more than doubled the number of chips, moved to a newer SP computer and built software tools for evaluation tuning and debugging [4]. They concluded that Deep Blue's searching ability was acceptable and spent most of their time between the matches designing, testing and tuning the new evaluation function [4]. Before the rematch, Deep Blue Jr. beat grandmasters Larry Christiansen and Michael Rohde, each by 1.5-0.5 [4].

How did Deep Blue beat Kasparov in 1997?

The rematch ran from May 3 to May 11, 1997, at the Equitable Center in New York City [2][7]. The winner was to receive $700,000 and the loser $400,000 [7][25]. Deep Blue won 3.5 to 2.5. The New York Times called it "the first time any chess champion has been beaten by a machine in a traditional match" [18], and the team received the Fredkin Prize for defeating the human world champion in a regulation match [4], a $100,000 award offered since 1980 [25].

Game-by-Game Summary

Game 1 (May 3): Kasparov wins. Kasparov, playing White, opened with a Réti that developed into the King's Indian Attack and won in 45 moves [7]. Late in the game, in an already lost position, Deep Blue played the odd 44...Rd1. Campbell later told Nate Silver that it was a bug. In Silver's account, "unable to select a move, the program had defaulted to a last-resort fail-safe in which it picked a play completely at random"; the bug was inconsequential and was fixed the next day [21]. "We had seen it once before, in a test game played earlier in 1997, and thought that it was fixed," Campbell told Silver [21]. Silver argued that the bug "was likely what allowed the computer to beat Kasparov," by feeding Kasparov's anxiety before game 2 [21]. In a 2014 NPR interview, Campbell said, "Due to a bug in the program, unfortunately, it had played a random move," and offered a theory that Kasparov, overestimating the machine, later missed a drawing chance, adding, "And I don't know if this is true or not" [22]. Mig Greengard, Kasparov's longtime aide and publicist, told NPR that Kasparov rejected the idea that the move affected his later play [22].

Game 2 (May 4): Deep Blue wins. The game was a Ruy Lopez, Smyslov Variation [7], and Kasparov later traced his problems in the match to it [18]. At move 36, Deep Blue declined to grab pawns with 36.Qb6 and instead played 36.axb5 axb5 37.Be4, a quiet move that grandmaster John Nunn called "A cruel move" in his annotations; Nunn noted that "most computers (including Fritz and Hiarcs) go for the material without hesitation" in the alternative line [8]. Kasparov resigned after 45 moves, but post-game analysis showed that the final position was drawn: after 45...Qe3, Black could have forced a perpetual check, a chance Deep Blue had allowed with 44.Kf1 [7][18]. Kasparov later said he had missed the draw because the computer had played so well that he assumed it could not allow one: "I do not understand how the most powerful chess machine in the world could not see simple perpetual check" [18]. Greengard summed up the game as Kasparov saw it: "How could something play like God, then play like an idiot in the same game?" [22] After this game, Kasparov suggested that a grandmaster had been behind one of Deep Blue's moves [7].

Game 3 (May 6): Draw. Kasparov opened with the irregular 1.d3, which transposed into an English Opening, hoping to take Deep Blue out of its opening book. The game was drawn in 48 moves [7].

Game 4 (May 7): Draw. The game began 1.e4 c6 and transposed into a Pirc Defense. Kasparov fell into time trouble late in the game, which ended in a draw after 56 moves [7]. This was the only game of the match that approached an endgame covered by the databases [4].

Game 5 (May 10): Draw. Kasparov repeated the Réti and King's Indian Attack of game 1. Deep Blue held a draw in an endgame that had looked likely to go Kasparov's way, and the game was drawn after 49 moves [7]. Analysis with the engine Komodo 8, published by ChessBase in 2015, found that Kasparov had missed a win starting with 44.Rg7+ [39]. Kasparov said that after game 5 he had become so dispirited that he felt the match was already over [18].

Game 6 (May 11): Deep Blue wins. With the score tied at 2.5-2.5, the final game decided the match.

GameDateResultMovesOpening
Game 1May 3Kasparov wins45Réti Opening (King's Indian Attack)
Game 2May 4Deep Blue wins45Ruy Lopez, Smyslov Variation
Game 3May 6Draw481.d3, transposing to the English Opening
Game 4May 7Draw561.e4 c6, transposing to the Pirc Defense
Game 5May 10Draw49Réti Opening (King's Indian Attack)
Game 6May 11Deep Blue wins19Caro-Kann Defense, Steinitz Variation
Final ScoreDeep Blue 3.5, Kasparov 2.5

Expanded article table

Why was Game 6 so controversial?

Game 6 was by far the shortest game of either match [7]. Kasparov, playing Black and needing a win, chose the Caro-Kann Defense, a conservative opening in which, The New York Times wrote, he was "perhaps too defiant in the early going, pursuing a risky sequence of moves" [18]. His 7...h6 has been described as a possible mix-up of move order, and Deep Blue answered with 8.Nxe6, a knight sacrifice well known in opening theory [9]. With his queen lost and his king exposed, Kasparov resigned after 19 moves [18], in a game reported to have lasted about an hour [9][10]. "I lost my fighting spirit," he said [18].

Observers disagreed about the final position. Grandmaster Patrick Wolff said, "His resignation was probably premature, but he was probably lost" [18]. Grandmaster John Fedorowicz, quoted by HISTORY, said, "Everybody was surprised that he resigned because it didn't seem lost" [10]. The Times noted that the defeat was unprecedented in Kasparov's career: he had never before lost a multigame match against an individual opponent [18].

At the news conference after the match, Kasparov traced his problems to the second game: "Game 2 had dramatic consequences, and I never recovered" [18]. He said he was frustrated that IBM would not let him see printouts of the computer's thought processes. Asked whether he was accusing IBM of cheating, he said, "I have no idea what's happening behind the curtain" [18]. He said any new match should not be sponsored by IBM, should last at least 10 games over 20 days, and should give him access to the computer's previous games [18].

Kasparov suggested that Deep Blue had been aided by some form of human assistance, charges IBM denied; he demanded a rematch, but IBM retired Deep Blue instead [6]. Team leader C. J. Tan said the computer logs would be published in appropriate journals [18]. IBM later released the complete logs of the games on its website; ChessBase's Albert Silver dates the release to 2000 [5].

How did Kasparov's view change?

Kasparov's view softened two decades later. In a December 2016 podcast with Sam Harris, reported by ChessBase, he said that while researching a new book he had analyzed the games with modern computers: "I changed my conclusions." He said his respect for the Deep Blue team had gone up and his opinion of his own play, and of Deep Blue's, had gone down [23]. In that book, Deep Thinking (2017), he wrote of the question of whether Deep Blue cheated: "After twenty years of soul-searching, revelations, and analysis, my answer is now 'no.'" [24] He still lamented that, after IBM declined a rematch, "the real victim of this betrayal was science," as reviewer Jonathan Schaeffer reported in MIT Technology Review [24].

What was Deep Blue's cultural impact?

Both matches drew heavy publicity [2][6]. During the 1996 match, IBM's website offered computer users a view of the games, and Carol Moore, director of IBM's homepage, said the company had expected no more than 200,000 users to visit the site in one day, but traffic topped 2 million on the first day [26]. Ahead of the 1997 rematch, NPR noted that the match would be available on the Internet as it happened [27].

The cultural impact went well beyond chess:

Public perception of AI. IBM describes the rematch as "widely seen as a symbolic test of whether supercomputers were catching up to human intelligence" [2]. Computer scientist Drew McDermott wrote in The New York Times days after the match that he had told his students, after Kasparov's 1996 win, "that it would be many years before computers could challenge the best humans," and that "Now that I and many others have been proved wrong, a lot of people have been rushing to assure us that Deep Blue is not actually intelligent" [19]. He argued that Deep Blue was "a 'little bit' intelligent," adding: "Saying that Deep Blue doesn't really think is like saying an airplane doesn't really fly because it doesn't flap its wings" [19].

The skeptical response. Other AI researchers were less impressed by the method. John McCarthy, who coined the term artificial intelligence, wrote in a 1997 review in Science that "it is a measure of our limited understanding of the principles of artificial intelligence (AI) that this level of play requires many millions of times as much computing as a human chess player does." Recalling Alexander Kronrod's description of chess as "the Drosophila of artificial intelligence," McCarthy added that "computer chess has developed much as genetics might have if the geneticists had concentrated their efforts starting in 1910 on breeding racing Drosophila. We would have some science, but mainly we would have very fast fruit flies" [20].

Chess community. The match showed that even the strongest human players could be beaten by a sufficiently powerful computer. IBM says the upgrades it made for 1997 also led to better training tools for chess players in later years [2]. Today the developers of Stockfish, an open-source engine, describe it as "stronger than any human, even when running on older or low-end hardware" [33], and AlphaZero's developers report superhuman play in chess, shogi and Go [11].

IBM's brand. HISTORY calls the 1997 victory "a huge publicity boost for IBM" [6].

The match was also the subject of the 2003 documentary film Game Over: Kasparov and the Machine, which repeated Kasparov's claim about game 2 [7].

Why did IBM retire Deep Blue?

After the 1997 victory, IBM declined Kasparov's request for a rematch and retired Deep Blue [6]. Campbell explained the decision in 2017: "We felt we had achieved our goal, to demonstrate that a computer could defeat the world chess champion in a match and that it was time to move on to other important research areas" [14].

The Deep Blue SP system consisted of two towers [29]. One of them, a gift of IBM, is in the collection of the National Museum of American History, part of the Smithsonian Institution in Washington, D.C.; it measures 79 x 36 x 47 inches, and the museum's record, as archived in 2024, listed it as "Currently not on view" [28]. The Computer History Museum in Mountain View, California, featured one of the two towers in "Mastering the Game," a computer-chess exhibit that ran from 2005 to 2012 and continues online [29].

Aftermath for IBM

IBM says that after Deep Blue it went on to develop new kinds of massively parallel computers, such as Blue Gene and Watson, and that in 1999 it built on its Deep Blue experience to launch the Deep Computing Institute [2]. The company also credits Deep Blue's architecture with later uses in financial modeling, data mining and drug development [2].

IBM's later AI showcase, Watson, played the television quiz show Jeopardy!. Campbell has contrasted it with Deep Blue: Watson used a machine-learning-based system that drew on large amounts of existing text, such as Wikipedia, to learn how to answer questions [14]. That approach relied on natural language processing and information retrieval rather than game-tree search.

Hsu published Behind Deep Blue: Building the Computer that Defeated the World Chess Champion in 2002, an insider's account of the project [1].

Deep Blue Team MemberRole on the projectLater affiliation
Feng-hsiung HsuStarted ChipTest at Carnegie Mellon; designed the chess chips [2][3]Compaq Computer Corporation at the time of the 2002 paper [4]; later Microsoft Research Asia [37]; Iridium Medical Technology as of 2017 [38]
Murray CampbellDeep Thought co-developer; "the AI expert," in his own words [14]IBM T. J. Watson Research Center, as a distinguished research staff member in 2017 [14]
A. Joseph Hoane Jr.Principal software engineer, responsible for the parallel search [36]Sandbridge Technologies at the time of the 2002 paper [4]
Joel BenjaminGrandmaster consultant; main author of the opening book [4][31]
C. J. TanTeam leader during the 1997 match [18]

Expanded article table

How does Deep Blue compare to modern chess AI?

Deep Blue's approach, built on massively parallel hardware and alpha-beta search, followed a line of research going back to Claude Shannon's 1950 paper "Programming a Computer for Playing Chess" [3]. Its methods, though, were specific to chess. Its 480 custom chips, its hand-designed evaluation function and its opening book were all built for chess, and the system could not play Go, understand language or recognize images. That narrow focus led some researchers, McCarthy among them, to question what Deep Blue taught about intelligence [20].

Evolution of Computer Chess After Deep Blue

In the years after Deep Blue's victory, custom chess hardware gave way to software running on ordinary computers, and engine strength kept rising [5].

YearEngine / SystemKey development
1997Deep BlueCustom chess chips and a 30-node parallel computer; beat Kasparov 3.5-2.5 [4]
2005-2006RybkaChessBase's Albert Silver credits Rybka's author, Vasik Rajlich, with introducing engine testing through tens of thousands of extremely quick games, and writes that engines had slowed to 30-40 Elo of progress per year, "if that," before Rybka, while afterward gains of 80-100 Elo were "no longer unheard of" [5]
2008StockfishOpen-source engine forked from version 2.1 of Glaurung, first announced on November 8, 2008 [35]
2017AlphaZeroSelf-play reinforcement learning with a neural network and Monte Carlo tree search; beat Stockfish 8 [11][15]
2018Leela Chess ZeroOpen-source, AlphaZero-style engine trained with distributed volunteer computing [13]
2020Stockfish 12NNUE neural-network evaluation enabled by default inside alpha-beta search [32]
2026Stockfish 18 and 19New NNUE architectures; Stockfish 19 reports an Elo gain of up to 44 points over Stockfish 18 in its own tests [33][34]

Expanded article table

The 1997 Deep Blue played only six games, all against Kasparov [3], so the project's strength estimates were indirect. Hsu estimated that Deep Blue Jr., the single-node version, played at "2700 plus" on the strength of its results against the project's grandmasters, and noted that Kasparov was rated around 2815 [3]. Modern engines are clearly stronger than the 1997 machine even on everyday hardware. ChessBase wrote in 2014 that Komodo 8 running on a desktop "or even on a modern quad-core smartphone" would "beat Deep Blue quite handily" [5], and Kasparov said in 2016, "Today you can buy a chess engine for your laptop that will beat Deep Blue quite easily" [23].

How does Deep Blue differ from AlphaZero?

In 2017, DeepMind's AlphaZero took a fundamentally different approach. DeepMind described it first in a December 2017 arXiv preprint and then in a peer-reviewed 2018 paper in Science, and the two versions report different numbers [11][15]. Instead of hand-crafted evaluation functions and specialized chess hardware, AlphaZero used deep reinforcement learning to teach itself chess, Go and shogi from random play, "given no domain knowledge except the game rules" [11]. Both versions report that it outperformed Stockfish after just 4 hours (300,000 steps) of training [11][15]. In the preprint's 100-game match at one minute per move against the 2016 TCEC world champion Stockfish (version 8), AlphaZero won 28 games, drew 72 and lost none [15]. In the Science version, after about 9 hours of training, it won 155 games and lost 6 out of 1,000 against Stockfish 8, at 3 hours per game plus 15 seconds per move [11].

Search volume shows the contrast most clearly. The preprint states that "AlphaZero searches just 80 thousand positions per second in chess and 40 thousand in shogi, compared to 70 million for Stockfish and 35 million for Elmo" [15]. The Science paper gives 60,000 positions per second for AlphaZero against 60 million for Stockfish in its main text, and 63,000 against 58.1 million in its supplementary Table S4 [11]. AlphaZero, the authors write, "may compensate for the lower number of evaluations by using its deep neural network to focus much more selectively on the most promising variations," using Monte Carlo tree search rather than alpha-beta search [11].

AlphaZero's playing style also stood out. DeepMind founder and CEO Demis Hassabis said at the NIPS conference in December 2017, "It's like chess from another dimension" [16]. Kasparov wrote in Science that "AlphaZero prioritizes piece activity over material, preferring positions that to my eye looked risky and aggressive," adding that he was pleased to see that it had "a dynamic, open style like my own" [17].

FeatureDeep Blue (1997)AlphaZero (2017 preprint, 2018 Science paper)Stockfish (NNUE era, 2020-2026)
SearchAlpha-beta search in software and on custom chips [4]Monte Carlo tree search guided by a deep neural network [11]Alpha-beta search with an NNUE neural network evaluation [32]
Chess knowledge suppliedAbout 8,000 hand-designed evaluation features, a grandmaster-prepared opening book and endgame databases [4]The game rules only [11]Network trained on the evaluations of millions of positions in Stockfish 12 [32]; "hundreds of billions of training positions," rescored using a strong Leela network, in Stockfish 19 [34]
Search speedAbout 100 to 200 million positions per second, depending on the position [4]80,000 positions per second (preprint) [15]; 60,000 (Science main text) [11]Hardware dependent; over 500 million positions per second on the highest-end hardware, per the Stockfish 18 release notes [33]
Hardware30-node IBM RS/6000 SP with 480 custom chess chips [4]Trained with 5,000 first-generation TPUs (Google's AI accelerator chips) plus second-generation TPUs (64 in the preprint, 16 in Science); played on a single machine with four first-generation TPUs [11][15]Standard CPUs [32]
GamesChess onlyChess, shogi and Go [11]Chess, including Fischer Random Chess [33]
LearningNone at play time; weights hand-tuned, with some automated tuning aids [4]Self-play reinforcement learning from random play [11]NNUE network trained offline [32][34]

Expanded article table

The contrast between Deep Blue and AlphaZero illustrates a broader shift in AI from hand-engineered, domain-specific systems to general-purpose learning systems. Deep Blue's knowledge came from features that its designers and grandmaster advisers specified and tuned [4]. AlphaZero learned its evaluation from self-play, starting with no knowledge beyond the rules [11].

Leela Chess Zero

After AlphaZero's publication, Stockfish co-author Gary Linscott announced Leela Chess Zero (Lc0) on January 9, 2018. The open-source project set out to reproduce AlphaZero's deep learning and Monte Carlo tree search techniques, using distributed volunteer computing to train the network's weights [13]. Lc0 won the Top Chess Engine Championship (TCEC) superfinal against Stockfish in Season 15 in May 2019 (53.5-46.5) and again in Season 17 in April 2020 (52.5-47.5); Stockfish won the Season 18 superfinal [13]. The two engines illustrate two approaches in modern computer chess: alpha-beta search with a neural network that can be evaluated efficiently on CPUs (Stockfish) and neural-network-guided Monte Carlo tree search (Lc0) [13][32]. The approaches have also mixed: Stockfish 19's training data was rescored using a Leela network [34].

Broader Significance

Deep Blue's victory over Kasparov is sometimes cited as a milestone marking the moment computers matched the best human performance in a domain long treated as a benchmark for intelligence. The AlphaZero authors call chess "the longest-studied domain in the history of artificial intelligence" [11]. Hsu placed building the first world-champion-class chess computer "among the oldest challenges in computer science" [3], and in 1965 the mathematician Alexander Kronrod called chess "the Drosophila of artificial intelligence" [20]. Deep Blue met that challenge through massive search, and McCarthy argued that "the fixation of most computer chess work on success in tournament play has come at scientific cost" [20].

The match also illustrated a recurring pattern in AI history: once a computer solves a problem, many people stop counting it as intelligence. McDermott observed this within days of the match, as commentators rushed to explain that Deep Blue was not really intelligent [19]. After Deep Blue, high-profile game challenges moved to harder targets, such as Go, where Google's AlphaGo beat Lee Sedol in 2016 [38], and AI research turned toward natural language understanding and more general reasoning. Within the longer history of AI, Deep Blue stands as a landmark of the symbolic, search-based approach that Campbell called "good old-fashioned AI" [14].

Deep Blue remains an important chapter in AI history, less for the specific techniques it used than for what it represented: in the words of The New York Times, "the first time any chess champion has been beaten by a machine in a traditional match" [18].

Continue reading

More to read

Topics:Artificial Intelligence

References

  1. ^Hsu, F.-h. (2002). *Behind Deep Blue: Building the Computer that Defeated the World Chess Champion.* Princeton University Press.
  2. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12 ^13 ^14 ^15 ^16"Deep Blue." IBM History. ibm.com/...deep-blue
  3. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12 ^13 ^14 ^15 ^16 ^17 ^18Hsu, F.-h. (1999). "IBM's Deep Blue Chess Grandmaster Chips." IEEE Micro, 19(2), 70-81. archive.computerhistory.org/...EE.062303055.sm.pdf
  4. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12 ^13 ^14 ^15 ^16 ^17 ^18 ^19 ^20 ^21 ^22 ^23 ^24 ^25 ^26 ^27 ^28 ^29 ^30 ^31 ^32 ^33 ^34 ^35 ^36 ^37 ^38 ^39 ^40 ^41 ^42 ^43 ^44 ^45 ^46 ^47 ^48 ^49 ^50 ^51 ^52 ^53 ^54 ^55 ^56 ^57 ^58 ^59 ^60 ^61 ^62 ^63 ^64 ^65 ^66 ^67 ^68 ^69 ^70 ^71 ^72 ^73Campbell, M., Hoane, A. J., Jr., & Hsu, F.-h. (2002). "Deep Blue." Artificial Intelligence, 134(1-2), 57-83. sciencedirect.com/...S0004370201001291
  5. ^1 ^2 ^3 ^4Silver, A. (December 26, 2014). "Komodo 8: Deep Blue revisited (part one)." ChessBase. en.chessbase.com/...8-deep-blue-revisited-part-one
  6. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9"Chess champion Garry Kasparov defeats IBM's Deep Blue." HISTORY. history.com/...arov-defeats-chess-playing-computer
  7. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12 ^13"Deep Blue versus Garry Kasparov." Wikipedia. en.wikipedia.org/...Deep_Blue_versus_Garry_Kasparov
  8. ^Silver, A. (February 19, 2015). "Deep Blue's cheating move." ChessBase. en.chessbase.com/...deep-blue-s-cheating-move
  9. ^1 ^2"Deep Blue versus Kasparov, 1997, Game 6." Wikipedia. en.wikipedia.org/..._versus_Kasparov,_1997,_Game_6
  10. ^1 ^2"Deep Blue defeats Garry Kasparov in chess match." HISTORY. history.com/...feats-garry-kasparov-in-chess-match
  11. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12 ^13 ^14 ^15 ^16Silver, D. et al. (2018). "A general reinforcement learning algorithm that masters chess, shogi, and Go through self-play." Science, 362(6419), 1140-1144. Full text: storage.googleapis.com/...alphazero_preprint.pdf
  12. ^"Thompson's Databases." Chessprogramming wiki. chessprogramming.org/Thompson%27s_Databases
  13. ^1 ^2 ^3 ^4"Leela Chess Zero." Chessprogramming wiki. chessprogramming.org/Leela_Chess_Zero
  14. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12 ^13Greenemeier, L. (June 2, 2017). "20 Years after Deep Blue: How AI Has Advanced Since Conquering Chess." Scientific American. scientificamerican.com/...d-since-conquering-chess
  15. ^1 ^2 ^3 ^4 ^5 ^6 ^7Silver, D. et al. (2017). "Mastering Chess and Shogi by Self-Play with a General Reinforcement Learning Algorithm." arXiv:1712.01815. arxiv.org/...1712.01815
  16. ^Knight, W. (December 8, 2017). "Alpha Zero's 'Alien' Chess Shows the Power, and the Peculiarity, of AI." MIT Technology Review. technologyreview.com/...-and-the-peculiarity-of-ai
  17. ^Kasparov, G. (2018). "Chess, a Drosophila of reasoning." Science, 362(6419), 1087. science.org/...science.aaw2221
  18. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12 ^13 ^14 ^15 ^16 ^17 ^18Weber, B. (May 12, 1997). "Swift and Slashing, Computer Topples Kasparov." The New York Times. nytimes.com/...-slashing-computer-topples-kasparov
  19. ^1 ^2 ^3McDermott, D. (May 14, 1997). "Yes, Computers Can Think." The New York Times, Section A, p. 21. nytimes.com/...yes-computers-can-think
  20. ^1 ^2 ^3 ^4McCarthy, J. (June 6, 1997). "AI as Sport." Review of *Kasparov versus Deep Blue: Computer Chess Comes of Age* by Monty Newborn. Science. Author's copy: www-formal.stanford.edu/...newborn
  21. ^1 ^2 ^3Silver, N. (2012). *The Signal and the Noise.* Penguin Press, p. 288. Excerpted in Plumer, B. (September 26, 2012), "Nate Silver's 'The Signal and the Noise'." The Washington Post. web.archive.org/...ilvers-the-signal-and-the-noise
  22. ^1 ^2 ^3Foo, S. (August 8, 2014). "Kasparov Vs. Deep Blue." Snap Judgment, NPR. npr.org/...kasparov-vs-deep-blue
  23. ^1 ^2Friedel, F. (December 29, 2016). "Kasparov on the future of Artificial Intelligence." ChessBase. en.chessbase.com/...ure-of-artificial-intelligence
  24. ^1 ^2Schaeffer, J. (July 10, 2017). "Kasparov Thinks Deeply about His Battle with a Machine." MIT Technology Review. technologyreview.com/...-his-battle-with-a-machine
  25. ^1 ^2Boyle, A. (June 16, 2004). "How prizes pushed progress." NBC News. nbcnews.com/...wbna5191763
  26. ^"Web surfers overrun IBM web site." CNN, February 11, 1996. web.archive.org/...chess
  27. ^1 ^2"Deeper Blue." NPR, May 2, 1997. npr.org/...deeper-blue
  28. ^"Deep Blue Supercomputer Tower." National Museum of American History, Smithsonian Institution, object 2002.0251.01. americanhistory.si.edu/...nmah_1005331
  29. ^1 ^2 ^3"Deep Blue II." Mastering the Game, Computer History Museum. computerhistory.org/...art-43305f13ef377
  30. ^"Embedding Software in Hardware." Mastering the Game, Computer History Museum. computerhistory.org/...edding-software-in-hardware
  31. ^1 ^2"Grandmaster Joel Benjamin and member of the 1997 Deep Blue team." Mastering the Game, Computer History Museum. computerhistory.org/...stl-431614f6804f8
  32. ^1 ^2 ^3 ^4 ^5 ^6"Stockfish 12." Stockfish blog, September 2, 2020. stockfishchess.org/...stockfish-12
  33. ^1 ^2 ^3 ^4"Stockfish 18." Stockfish release notes, January 31, 2026. github.com/...sf_18
  34. ^1 ^2 ^3 ^4"Stockfish 19." Stockfish release notes, September 5, 2026. github.com/...sf_19
  35. ^"Stockfish." Chessprogramming wiki. chessprogramming.org/Stockfish
  36. ^1 ^2"Joe Hoane." Chessprogramming wiki. chessprogramming.org/Joe_Hoane
  37. ^"Feng-hsiung Hsu." Chessprogramming wiki. chessprogramming.org/Feng-hsiung_Hsu
  38. ^1 ^2"Chip Hall of Fame: IBM Deep Blue 2 Chess Chip." IEEE Spectrum, June 30, 2017. spectrum.ieee.org/...me-ibm-deep-blue-2-chess-chip
  39. ^Silver, A. (January 9, 2015). "Komodo 8: Deep Blue revisited (part three)." ChessBase. en.chessbase.com/...deep-blue-revisited-part-three

Improve this article

Add missing citations, update stale details, or write a clearer explanation. Anyone can edit: changes go live when you publish, and every edit is checked by an editor.

7 revisions · v7 · 7,025 words · full history

Fact-checks are independent of edits: a reviewer re-verifies the article against its sources and stamps the date. How we verify

Research and drafting on this wiki are AI-assisted, under named human editorial standards. How AI is used here

Reviewer note: Fact-checked on 2026-10-30 against Campbell, Hoane and Hsu (Artificial Intelligence, 2002), IBM and museum records, contemporaneous New York Times reporting and game records, and later engine release notes, with an independent second review.

Cite this page: AI Wiki. "Deep Blue." aiwiki.ai, updated 30 Oct 2026, fact-checked 30 Oct 2026. CC BY 4.0. https://aiwiki.ai/wiki/deep_blue

Edit article