For each of these defeats the year is known and so are the names. The other half is quoted far less often: how much hardware stood on the machine's side. Without it, fourteen facts about games quietly turn into one claim about people.
Every row here carries three hardware columns: what was in the match itself, what an ordinary person had on their desk that same year, and what does the same work now. The human side is identical in all fourteen rows and needs no column of its own: one brain, about twenty watts.
The word "price" in the title is a figure of speech: what is counted here is processors, nodes and positions, not money. What the same machines cost in dollars is on the neighbouring page, What Winning Cost. What they drew from the wall while doing it is in The Watt League.
The rank correlation between the size of a game and the year it fell, across the nine games whose size is defined at all: ρ = 0.55. At n = 9 it would need to reach 0.68 to mean anything. So the connection is there to the eye and absent in the arithmetic — and five of the fourteen games cannot be put on this axis at all, all five of them fallen since 2011.
Tree sizes come from the Solved Games page and the site's own cards; dates from the How Many Days Ago page. Spearman's rank correlation is recomputed on this page at every load, from the same data the dots are drawn from. The two-sided 5% critical value at n = 9 is 0.683 (Zar, Biostatistical Analysis, table B.20).
Hans Berliner's BKG 9.8 beats world champion Luigi Villa 7-1 in Monte Carlo
One PDP-10 in Pittsburgh. The program never travelled to the tournament — the moves went over a satellite link.
The MOS 6502 out of an Apple II: eight bits in a forty-pin package. Between it and a PDP-10 lay essentially everything — the price, the size, the room, the staff. in the catalogue
A free neural-network bot on a phone plays stronger than BKG 9.8 and needs no satellite time.
H. Berliner, Scientific American 242(6), 1980
James Allen and Victor Allis solve the game independently: with perfect play the first player wins
No figure exists. Both were university projects, and neither Allen nor Allis published what they ran on. What is on record is that no full enumeration was needed: Allis's database of roughly half a million rules and positions did the work instead.
The Intel 80386 of 1985 — thirty-two bits for those who could afford it. Most desks in 1988 still had a 286, some an XT. It is the nearest chip to the year in the catalogue. in the catalogue
The whole solution fits in a web page; a browser answers any position instantly.
L. V. Allis, A Knowledge-based Approach of Connect-Four, MSc thesis, Vrije Universiteit Amsterdam, 1988; J. D. Allen, The Complete Book of Connect 4, 2010
Ralph Gasser proves that perfect play is a draw
There is no single number to give. The computation ran for four years on whatever was available: the first database on a Macintosh IIx in 1989, the last on a Macintosh Quadra 800 in 1993, and in between a Cray X-MP/28, a VAX 9000-420, an IBM RS/6000, a sixteen-transputer rig and a thirty-processor MUSIC machine.
An Intel 80486 — the ordinary desk machine of 1993. The machine side of this row was very nearly its equal: the last database was finished on a desktop Macintosh. in the catalogue
The complete Morris database is an ordinary file on an ordinary desktop drive.
R. Gasser, Solving Nine Men's Morris, Computational Intelligence 12(1), 1996; R. Gasser, Harnessing Computational Resources for Efficient Exhaustive Search, PhD thesis, ETH Zürich, 1995
Victor Allis solves gomoku: the first player wins, however hard the human tries as second
What was published is machine time, not a machine: on the order of 1.1 to 1.3 million CPU-seconds per variant — weeks of computing. How many processors, and of what kind, the paper does not say, so this row has no bar.
The same 486 as the row above. Two falls in one year, and the trees behind them differ by ninety-five orders of magnitude: no argument against "size decides" on this page is plainer than that. in the catalogue
The full traversal was never needed. The proof-number search invented for that job is now a textbook technique.
L. V. Allis, H. J. van den Herik, M. P. H. Huntjens, Go-Moku Solved by New Search Techniques, AAAI Fall Symposium, 1993; L. V. Allis, Searching for Solutions in Games and Artificial Intelligence, PhD thesis, 1994
Chinook takes the world title: Marion Tinsley withdraws for health reasons after six draws
An eighteen-processor Silicon Graphics Challenge with a gigabyte of memory and the endgame databases on disk.
An Intel Pentium, on sale since the year before: one processor at 60 to 66 MHz. Eighteen processors in one cabinet was not a consumer product in any sense. in the catalogue
An engine of that strength is a free download and runs on a single core.
Chinook project archive, University of Alberta; J. Schaeffer, One Jump Ahead, 1997
Deep Blue beats Kasparov 3½-2½ in New York; the decisive sixth game lasts nineteen moves
Thirty RS/6000 SP nodes driving 480 chess chips fabricated for this one task — about two hundred million positions a second.
The Intel Pentium II went on sale on 7 May 1997, four days before game six. One processor, 233 MHz. in the catalogue
Deep Blue stayed the last machine of the old kind: today's programs search orders of magnitude fewer positions and play better.
F.-h. Hsu, IEEE Micro 19(2), 1999; M. Campbell, A. J. Hoane, F.-h. Hsu, Artificial Intelligence 134(1-2), 2002
Michael Buro's Logistello beats world champion Takeshi Murakami six games to nil in Princeton, 4-7 August
Buro never published the match configuration. Logistello was a single-processor program: the speeds he quoted were for a Pentium Pro 200 (about 160,000 positions a second in the middle game) and a DEC Alpha 500. The bar shows one processor, on the same reasoning the matches page applies to KataGo.
The AMD K6 arrived that April — the same class of chip as the one Logistello ran on. The gap between the machine side of this table and the desk side is almost nil, and almost nobody covered the match. in the catalogue
The strongest open-source Othello programs run on a single core.
M. Buro, The Othello Match of the Year: Takeshi Murakami vs. Logistello, ICCA Journal 20(3), 1997; M. Buro, The Evolution of Strong Othello Programs, 2002
Jonathan Schaeffer's team proves perfect play is a draw. The game closes — not for humans, for everyone
Eighteen years of computation begun in 1989: an average of about fifty machines a day, with more than two hundred at peak. The bar shows the average, because the average is what the eighteen years measured.
The Intel Core 2 Quad Q6600 — the first affordable four cores on a desk. Over the same years the computation ran, the desk machine went from a 486 to this. in the catalogue
The proof databases are public: any position's value can be checked from a browser.
J. Schaeffer et al., Checkers Is Solved, Science 317(5844), 2007
IBM Watson beats Ken Jennings and Brad Rutter at a game built on puns and hints
Ninety IBM Power 750 servers, 2,880 POWER7 cores, 16 TB of RAM — in a room next to the studio, with no internet access at all during the show.
The Intel Core i5-2500K shipped a month before the broadcast: four cores at 3.3 GHz. Three orders of magnitude below the other side of the table. in the catalogue
The same job is done by a language model that fits on one graphics card.
D. Ferrucci et al., Building Watson: An Overview of the DeepQA Project, AI Magazine 31(3), 2010
AlphaGo beats Lee Sedol 4-1 in Seoul. The human won game four; to this day it remains the last one
Forty-eight first-generation tensor processing units spread across several machines in Google Cloud. The figure is DeepMind's own, but it appeared only eighteen months after the match, in the AlphaGo Zero paper. The much-quoted 1,920 CPUs and 280 GPUs is the largest line of a scalability table in the 2016 paper and was never the Seoul machine. How many machines served those 48 TPUs was not published.
A four-core Intel. The nearest catalogue entry is the 2013 Haswell, and that is not a gap in the catalogue: Intel's desktop line had barely moved for three years by 2016. in the catalogue
Open-source KataGo plays at that level on roughly one graphics card.
D. Silver et al., Mastering the game of Go with deep neural networks and tree search, Nature 529(7587), 2016; the 48 TPUs are stated in D. Silver et al., Mastering the game of Go without human knowledge, Nature 550(7676), 2017
Over twenty days and 120,000 hands Libratus beats four professionals at no-limit hold'em
One hundred processors while playing. Overnight the bot took about six hundred nodes of the Bridges supercomputer and rewrote its own strategy: some fifteen million core-hours in total.
The AMD Ryzen 7 1800X went on sale a month after the match: eight cores on a desk. Whatever distance the machine side covered in forty years, the desk side covered too. in the catalogue
Hold'em solvers are sold as ordinary software and run on a desktop.
N. Brown, T. Sandholm, Superhuman AI for heads-up no-limit poker: Libratus beats top professionals, Science 359(6374), 2018; the hundred-processor figure is the comparison given in N. Brown, T. Sandholm, Science 365(6456), 2019
Pluribus beats five professionals at once, in a game with no tidy equilibrium theory to lean on
Two Intel Haswell E5-2695 v3 server processors and under 128 GB of memory while playing. The whole preparation was eight days on a sixty-four-core server: 12,400 core-hours.
The AMD Ryzen 5 3600 shipped four days before the paper: six cores for two hundred dollars. For the first time the machine side of the table cost less than the desk side. in the catalogue
Here the present arrived on the day of the result itself: two processors is a desktop computer.
N. Brown, T. Sandholm, Superhuman AI for multiplayer poker, Science 365(6456), 2019
AlphaStar reaches grandmaster level; OpenAI Five beats world champions OG — both results carrying caveats about restrictions
Two results and two figures that will not combine. OpenAI Five trained on 1,536 GPUs and 172,800 CPU cores; the AlphaStar league trained for fourteen days on sixteen third-generation TPUs per agent, and the trained agent then played from a single desktop GPU. Training and playing are different quantities, and putting them in one bar would be dishonest.
The Intel Core i9-9900K: eight cores at five gigahertz — the same gaming machine the humans on the other side were playing on. in the catalogue
The trained agents of both projects play from one GPU. The expensive part was the training, not the play.
O. Vinyals et al., Grandmaster level in StarCraft II using multi-agent reinforcement learning, Nature 575(7782), 2019; OpenAI, Dota 2 with Large Scale Deep Reinforcement Learning, arXiv:1912.06680, 2019
Sony AI's Sophy beats the game's best drivers, racing wheel to wheel without breaking racing etiquette
The DART training platform had over a thousand PlayStation 4 consoles available, with one NVIDIA V100 or half an A100 per training node; a single training run took ten to twenty consoles. The configuration of the races themselves was never published in comparable units, so there is no bar.
The Intel Core i9-12900K shipped three months before the paper: sixteen cores of two different kinds in one package. in the catalogue
Sophy is built into Gran Turismo 7 — you can race it on a console at home.
P. R. Wurman et al., Outracing champion Gran Turismo drivers with deep reinforcement learning, Nature 602(7896), 2022
The obvious explanation is that the hardware curve reached the size of the game and the game fell. It is a good story and the data does not support it. Backgammon and checkers are games of the same size, and fifteen years separate their falls. Gomoku is sixty-one orders larger than chess and fell four years earlier. Go is a hundred and twenty-six orders larger than chess and lagged it by nineteen years.
The site says the same thing from the other direction: How Machines Play lists six steps, and every one of them begins by giving up something the previous step treated as compulsory — while the horsepower grew steadily throughout. Deep Blue searched about two hundred million positions a second and stayed the last machine of the old kind; AlphaZero searches roughly a thousand times fewer and plays better. The hardware here is the price, not the cause. Which is exactly why it is worth counting: the price is the part nobody quotes.
The write-ups — How Machines Play, Five Matches, Solved Games — are in the games section. The hardware catalogue is in the history of processors.