← Back to the front page
The watt league
Energy as the third axis: what the machine drew, what a person draws, and why the two numbers are so hard to set side by side

There is a side to the machine's victories that gets discussed far less than the algorithm. A person at a board runs on roughly a hundred watts and cannot spend more: the ceiling is physiological and has not moved since 1979. The machine side, over the same forty-four years, went from a PDP-10 cabinet to ninety servers and back down to two processors. The thermodynamic mismatch is real and worth showing.

Showing it honestly is harder than it looks. The usual comparison is built the same way every time: twenty watts for the human brain against a whole machine room including its cooling, and sometimes against the energy of training, which the game that was actually played never touched. The ratio comes out in the tens of thousands and measures nothing, because both sides were chosen to make it large. Twenty watts of brain against watts at the wall is not a comparison, it is a choice of answer.

So every figure here carries the boundary it was taken at, a confidence tag and the derivation that produced it, and the human side is drawn as three lines rather than one. Training energy is never added to match energy and never stands beside it unmarked. And 8 of the 22 quantities have no number at all: in their place stands an account of what is missing and why. That is a finding too.

The three human lines
None of them is the right one. Which is right depends on what stands on the other side of the table
the brain at restthe organ alone20 WRpoint

The adult brain at rest. A physiological floor, not the price of a single move.

how it was derived

The twenty watts is a conversion, not a meter reading, and two independent routes arrive at it. One: Raichle and Gusnard (PNAS, 2002) give the brain as roughly 2% of body weight and roughly 20% of the body's oxygen and calorie consumption; 20% of an adult's basal rate (see body_resting, 60–90 W) is 12–18 W, and 20% of a seated person's rate (100–125 W) is 20–25 W. The other route is more direct: the classic Kety–Schmidt measurements put the brain's oxygen consumption at 3.3–4.0 ml per 100 g per minute (Sokoloff et al., 1955). For a 1,400 g brain that is 46–56 ml of oxygen a minute, and at an energy equivalent of about 20.9 J per ml it comes to 16–20 W. Both routes land on 16–21 W, and 20 is the rounding up. This does not earn an M: oxygen was measured, watts were not.

Sources: M. E. Raichle, D. A. Gusnard, Appraising the brain's energy budget (the brain is about 2% of body mass and takes about 20% of the oxygen and the calories the body uses; the rate is remarkably constant across widely varying mental and motor activity), 2002 arch. · L. Sokoloff, R. Mangold, R. L. Wechsler, C. Kennedy, S. S. Kety, The Effect of Mental Arithmetic on Cerebral Circulation and Metabolism (13 subjects; doing arithmetic in the head shifts the EEG and raises pulse and blood pressure, while cerebral blood flow and the brain's oxygen consumption do not move; control CMRO₂ 3.3–4.0 ml O₂ per 100 g per minute), 1955 arch.

the body at restthe whole body, basal rate60–90 WRrange

The whole body at rest — basal metabolic rate. An adult lying down, fasted, thermally comfortable.

how it was derived

Converted from kilocalories per day. One kcal/day = 4,184 J / 86,400 s = 0.04843 W. The Mifflin–St Jeor equation (1990), now the working standard in place of Harris–Benedict: a man of 70 kg, 175 cm, 40 years gives 1,599 kcal/day, which is 77 W; a man of 80 kg, 180 cm, 30 years gives 1,780 kcal/day, 86 W; a woman of 60 kg, 162 cm, 40 years gives 1,252 kcal/day, 61 W. Hence 60–90 W, covering ordinary adults of either sex. This is lying down and fasted; a person sitting at a board spends more — see the next row.

Sources: M. D. Mifflin, S. T. St Jeor, L. A. Hill, B. J. Scott, S. A. Daugherty, Y. O. Koh, A new predictive equation for resting energy expenditure in healthy individuals (the Mifflin–St Jeor equation), 1990 arch.

a person at the boardthe whole body, seated105–126 WRrange

A person sitting at a board. If watts are going to be set against watts, this is the honest figure to use.

how it was derived

From the metabolic heat production tables underlying the thermal comfort standards ISO 8996 and ASHRAE 55. The 'met' is defined as 58.2 W per square metre of body surface and describes a person seated at rest in thermal neutrality. The standard adult is 70 kg, 175 cm, with a DuBois surface area of 1.8 m². That gives: seated quiet, 1.0 met — 105 W; seated at office work, 1.2 met — 126 W. This is the body's total heat output, not the brain's share. A caveat: the ISO and ASHRAE tables descend from Passmore and Durnin's 1967 review and have not been revisited since. The spread among advisers — '70–100' against '100–120' — is not a disagreement but two different definitions: basal rate in the first case, a seated person in the second. '150–300 W' for someone sitting is simply wrong: on the same tables 150–210 W is standing and light physical work, and 300 W is walking.

Sources: International Organization for Standardization, ISO 8996: Ergonomics of the thermal environment — Determination of metabolic rate (tables of metabolic heat production by activity; 1 met = 58.2 W/m²), 2021 arch. · ASHRAE, ANSI/ASHRAE Standard 55: Thermal Environmental Conditions for Human Occupancy (the table of metabolic rates: seated quiet 1.0 met, seated office work 1.2 met; the standard adult's DuBois surface area 1.8 m²), 2023 arch.

the increment from mental workthe whole brain, task-evokedno dataO?

The increment from hard mental work. There is no reliable figure, and that is itself the finding: the increment is small enough to be hard to measure at all.

how it was derived

The most direct test was done seventy years ago and has not been overturned. Sokoloff, Mangold, Wechsler, Kennedy and Kety (Journal of Clinical Investigation, 1955) measured cerebral blood flow and oxygen consumption in 13 people at rest and while doing mental arithmetic. The EEG changed unmistakably during the arithmetic — a shift to low-voltage, high-frequency activity — and pulse and blood pressure rose, while the brain's oxygen consumption did not move: in their words, the control and arithmetic values were identical. The authors state the conclusion worth remembering the paper for: neither the degree of mental activity nor the EEG pattern correlates with total cerebral oxygen consumption.

Modern imaging did not overturn this so much as sharpen it. Locally, blood flow in a working region rises a great deal; across the whole brain, the task-evoked increase is usually under 5% of the ongoing baseline (Raichle and Mintun, Annual Review of Neuroscience, 2006). Five per cent of 20 W is under a watt, and that is a ceiling rather than a reading.

What is not here and will not be: the claim that a grandmaster burns 6,000 kilocalories a day. It comes from an ESPN piece of 2019 quoting Robert Sapolsky, and traces back to a tripled breathing rate multiplied by 2,000 kcal. No calorimeter appears anywhere in that chain. What does rise during a hard game is sympathetic arousal — heart rate, blood pressure, breathing, muscle tone. That raises the whole body's expenditure, not the brain's, and how much it raises it over a chess match has never been published.

Sources: L. Sokoloff, R. Mangold, R. L. Wechsler, C. Kennedy, S. S. Kety, The Effect of Mental Arithmetic on Cerebral Circulation and Metabolism (13 subjects; doing arithmetic in the head shifts the EEG and raises pulse and blood pressure, while cerebral blood flow and the brain's oxygen consumption do not move; control CMRO₂ 3.3–4.0 ml O₂ per 100 g per minute), 1955 arch. · M. E. Raichle, M. A. Mintun, Brain Work and Brain Imaging (the increase in the brain's energy use that a task brings about is small next to the baseline — usually under 5%), 2006 arch. · Aishwarya Kumar, The grandmaster diet: How to lose weight while barely moving (where the “6,000 calories a day” figure comes from, credited to Robert Sapolsky; no calorimetry underneath it), 2019 arch. · Andrew Gelman, About that bogus claim that “chess grandmasters” burn 6000 calories per day (where the number came from: respiration measured on 11 ordinary players, the highest reading passed off as the average, multiplied against 2,000 kcal), 2025 arch.

which source belongs to this number

For the statement 'the brain runs on about 20 watts', the correct citation is Raichle and Gusnard, PNAS 99(16), 2002. Their commentary says exactly what is needed: the brain is roughly 2% of body weight and accounts for roughly 20% of the body's oxygen and calories, and that rate is remarkably constant across widely varying mental and motor activity. The twenty watts follows from that share.

Attwell and Laughlin (Journal of Cerebral Blood Flow & Metabolism, 21, 2001) answer a different question — not how much, but on what. Theirs is an item-by-item budget of signalling in rodent grey matter, built bottom-up from anatomy and physiology: action potentials take 47% of the ATP, postsynaptic effects of glutamate 34%, maintaining the resting potential 13%, glutamate recycling 3%. No humans, no watts, and no whole brain appear in it. Citing it under 'the brain = 20 W' is a substitution: the paper is about the split, not the total, and not about our species.

Verdict: 20 W belongs to Raichle and Gusnard 2002. Keep Attwell and Laughlin 2001 where the subject is where the energy goes — there it is the primary source and irreplaceable. The brain-20-watts card in cards.js is attributed wrongly.

On any chart the human side needs at least two lines: the brain (20 W) and the person at the board (105–126 W). A single line against the machine's watts is not a comparison, it is a choice of answer. And neither line is truly commensurable with the machine's: the machine's number excludes its training, and the human's excludes the twenty years it took to learn the game.

The confidence scale
Four levels, one letter each; a trailing question mark means the letter itself is in doubt
M  measured

A direct measurement of this quantity, with a published method. If the source will not say where the meter was, it is not an M.

P  specification

The maker's own published figure: a TDP, a price list, a datasheet. Stated, but never measured in the installation described.

R  reconstructed

Derived from the configuration: a TDP times a count, the length of a game, a comparable machine. The derivation is in the method field.

O  order of magnitude

Good for scale only -- ten times, a hundred times -- and not for setting two numbers against each other.

On the chart the difference is visible without opening anything: a reconstruction is a solid mark with end caps, an order of magnitude is a pale band without them, and a tag carrying a question mark is outlined in a dashed stroke. Only a record holding a single number is drawn as a point; everything else is a band from end to end.

Watts by system
Nine systems in year order across, watts on a logarithmic scale up. Three panels rather than one: what the machine drew during play, what training cost, and what nobody ever published
10 W100 W1 kW10 kW100 kW1 MW1 kW10 kW100 kW1 MWperson at the board · 105–126 Wbody at rest · 60–90 Wbrain · 20 WIN PLAYIN TRAININGa different quantity — it does not add to the panel aboveNO NUMBERBKG 9.8 · compute nodes · 4.3–19.8 kW · ONBKG 9.8 · at the wall · 6.7–30 kW · O?WBKG 9.81979Chinook · at the wall · 5,700 W · PWChinook · at the wall · 3,200–5,700 W · RWChinook1994Deep Blue · chips only · 1,380 W · RCDeep Blue · at the wall · no data · O?WDeep Blue1997Logistello · chips only · 35 W · PCLogistello · at the wall · 100–200 W · OWLogistello1997Watson · compute nodes · 175.5 kW · RNWatson · the facility · no data · O?FWatson2011AlphaGo · chips only · 1,920 W · RCAlphaGo · compute nodes · 4.6–10.3 kW · ONAlphaGo · at the wall · no data · O?WAlphaGo · training · no data · O?TAlphaGo2016Libratus · the match · 12–144 kW · RMLibratus · training · 47 kW · RTLibratus · at the wall · no data · O?WLibratus2017Pluribus · the match · 240 W · RMPluribus · training · no data · O?TPluribus2019KataGo · training · 8.4–13.8 kW · RTKataGo · the match · no data · O?MKataGo · training · no data · OTKataGo2023
chips onlyC · where the boundary is
compute nodesN · where the boundary is
at the wallW · where the boundary is
the facilityF · where the boundary is
the matchM · which time is counted
trainingT · which time is counted
a person20 / 60–90 / 105–126 W

Every figure and every boundary comes from the site's system records; the derivation of each is printed under it in the section below. The horizontal axis is not linear: it is nine columns in year order, not a time scale. The upper and middle panels share one scale — a decade is the same height in both — but they measure different quantities and must not be added.

What exactly was counted
A watt without a boundary is not a quantity

Every watt figure raises two questions and both have to be answered. First, where the boundary was drawn: around the silicon, around the servers, at the wall socket, or around the whole room with its cooling. Between the extremes lies a factor of several, and the figure quoted is usually whichever one suits the speaker. Second, which time was counted: the game, or the training. That difference runs to orders of magnitude.

The dataset uses six scope values, and each record carries exactly the one a source could be pinned to. The empty places cluster at the wall and at the facility — the two boundaries people name out loud most often.

where the boundary is

chips onlyC3 records, 0 with no number

The silicon's dissipation: processors and accelerators and nothing else. No memory, no disks, no network, no power-supply losses, no cooling.

compute nodesN3 records, 0 with no number

Whole servers — silicon together with memory, disks and power supplies — but nothing of what stands around them in the room.

at the wallW7 records, 3 with no number

What the installation draws from the mains, all of it. Almost none of these exist here: it is the quantity everyone quotes and nobody published.

the facilityF1 record, 1 with no number

The installation together with its cooling, its uninterruptible supplies and everything else the room runs on its behalf.

which time is counted

the matchM3 records, 1 with no number

The hardware occupied by play while play is going on. Here that is the nameplate power of the processors doing the thinking: a narrow boundary over a short time.

trainingT5 records, 3 with no number

What it took to make the system able to play at all — before the match and outside it. It cannot be added to the match figure: training happened once, a game happens every time, and they are orders of magnitude apart.

Joules per hand
The one place in the dataset where both sides land in the same units

Nearly every machine figure here is a watt, which is power rather than energy: to reach joules you need to know how long the machine was thinking, and almost nowhere is that published. Pluribus is the exception. Its processors' power is known, its average pace is known, and the record carries a finished joules-per-hand figure. Dividing that by the power recovers the time it was built from — and the same time can be applied to a person.

the machine: 240 W × 20 s = 4,800 J R
a person at the board: 105–126 W × 20 s = 2,100–2,520 J R
a human brain: 20 W × 20 s = 400 J R
ratio against the person at the board: 1.9–2.3×
ratio against the brain alone: 12×

Two ratios out of the same machine joules: one against a person at the board, one against a brain alone. Five- or sixfold separates them, and that gap is the price of choosing a human line. No physics changed between the two rows. Only what was counted on the human side.

And even this comparison is incomplete on both sides. The machine's 240 watts is the nameplate figure for two processors, with no memory, no disk and no power-supply losses; the real wall draw was higher. The human's 105–126 watts is the whole body's heat output, not the brain's work. The machine's joules exclude its training and the human's exclude the years spent learning to play poker. The number above is honest exactly within the boundary it was computed in, and not one step beyond it.

The seconds are not typed in but recovered by division: the Pluribus record carries both the power and the joules per hand, and their quotient is the pace. The human watts come from the rows above — the person at the board and the brain at rest — carrying their own confidence tags.

Sources: N. Brown, T. Sandholm, Superhuman AI for multiplayer poker (p. 887: the blueprint strategy — 8 days on a 64-core server, 12,400 core-hours, under 512 GB of memory, “about $144 to produce” at spot cloud prices; play ran on two Intel Haswell E5-2695 v3 and under 128 GB; subgame search took 1 to 33 s, averaging 20 s per hand against copies of itself), 2019 arch. · Intel, Intel Xeon Processor E5-2695 v3 (35M Cache, 2.30 GHz) — Product Specifications: 14 cores, 28 threads, 2.30 GHz base, 3.30 GHz turbo, 22 nm, 120 W TDP, launched Q3 2014, 2014 arch.

Five cases
Where the comparison breaks, specifically

Deep Blue: the 6.8 kilowatts with no origin

1997 · chess · chips only 1,380 W · at the wall — no number

'Deep Blue drew 6.8 kilowatts' is easy to find in dozens of places and in no primary source. It is not in Hsu, who designed the chess chip, not in Campbell and colleagues, not in BYTE of 1997, not on IBM's own pages. An exact-string search returns no origin — and a figure with no origin cannot be labelled measured under any circumstances.

What is firm: the chip dissipated about a watt, there were 480 of them, and thirty node processors at thirty watts each. The addition gives 1,380 W of silicon, and that is a floor, not a ceiling: no memory, no disks, no SP switch, no supply losses, no cooling. What the installation pulled from the mains was never published by anyone. The empty cell in the 1997 column is the answer, and it is more accurate than any plausible number would be.

AlphaGo: a megawatt computed from the wrong machine

2016 · go · chips only 1,920 W · nodes 4.6–10.3 kW · at the wall — no number

'AlphaGo a megawatt, Lee Sedol twenty watts' is the most quoted energy comparison in the subject. It was born in a blog post written during the match: the author took the 1,920 processors and 280 graphics cards from the scalability table of the 2016 paper, allowed two hundred watts apiece and arrived at roughly a megawatt. He cites no source for the two hundred watts.

That is not the main problem. That configuration never played Lee Sedol: Seoul ran on 48 first-generation tensor processors. At the measured busy power of a die, that is 1,920 W of silicon, and 4.6–10.3 kW counting the servers holding them. The megawatt is an estimate of a different machine, and the twenty watts is the lowest of three human lines. 'Fifty thousand times' is the product of the two most flattering ends.

Libratus and Pluribus: the curve that went down

2017 · poker · in play 12–144 kW · training 47 kW → 2019 · poker · in play 240 W

Two systems from one lab and one author, two wins over professional poker players, two years apart. In 2017 Libratus occupied some six hundred nodes of a supercomputer: 12–144 kW counting processors alone, plus a separately reckoned strategy 'blueprint' at 47 kW, running for weeks.

In 2019 Pluribus played on two processors — 240 W — and its training fitted into one server over eight days. In core hours the gap is starker still: about six million for Libratus's blueprint alone against twelve thousand four hundred for Pluribus. The machine curve in this subject does not only rise, and any argument built on its rise owes this step an explanation.

Watson: a nameplate ceiling multiplied by ninety

2011 · quiz show · nodes 175.5 kW · the facility — no number

175.5 kW is the largest figure on this page, and it is a reconstruction rather than a meter reading. IBM publishes a maximum draw for one Power 750 Express server in its operating environment table; that is multiplied by 90 servers. It is the ceiling for a fully loaded machine: the actual load during the game was lower, and IBM never said by how much.

And this is still not the facility. Room cooling, uninterruptible supplies and external disk enclosures are outside the number, and no reading of the whole installation was published by IBM or by anyone watching. The 80 kW and 200 kW estimates in circulation do not lead back to a checkable source and are not repeated here.

KataGo: the human won and there are no watts

2023 · go · in play — no number · training 8.4–13.8 kW

The one case here where the human won is also the one where nothing is known about the machine's side of the game. The games ran on the public KGS server, on other people's machines; what hardware turned the program is stated neither in the paper nor on the project site. The only measurable quantity on the machine side is a hundred thousand visits per move, and that is work, not watts. Filling the gap with a plausible graphics card would be inventing a number.

What is known is the price of training, and it belongs to what came before the game rather than to the game: 8.4–13.8 kW of accelerators while KataGo itself was learning, and separately the training of the attacking policy — the very strategy the human then won with. The only watts that can be named beside a human victory are the price of finding a strategy, not the price of playing.

Every record
Nine systems, each quantity with its boundary, its tag and its derivation
BKG 9.81979 · BKG 9.8 vs Luigi Villa1 nodes

A PDP-10 at Carnegie-Mellon, linked by satellite to the hall in Monte Carlo, where a robot called Gammonoid displayed the moves. No source names the processor model: DEC serial-number lists credit the university with both KA10 machines (CMU-B, CMU-D) and KL10 machines (CMU-A, a DECsystem-1080; CMU-C, a KL2060). The machine was time-shared, so the game used a fraction of it rather than the whole installation. P?

compute nodes4.3–19.8 kWOrange

The bounds are PDP-10 processor-cabinet figures taken from DEC documentation and collated in a community discussion: roughly 4,300 W for a KA10 and roughly 19,800 W for a KL10. Which CMU machine actually played the match could not be established, so this is a spread across the class rather than a figure for the installation used.

Sources: members of the pidp-10 group, the figures collated from DEC documentation on bitsavers, Crowdfunding the PiDP-10 (and its size issues) — a thread on what a PDP-10 drew arch.

at the wall6.7–30 kWO?range

The lower bound is a KA10 processor cabinet (4,300 W) plus one MH10 memory module (2,400 W). The upper bound is the same discussion's estimate for a minimal KL10 system — processor, front end, one memory module and one disk drive — at about 30 kW. Room cooling is excluded. None of this can be charged to a single backgammon game: the machine was time-shared and the share is unknown.

Sources: members of the pidp-10 group, the figures collated from DEC documentation on bitsavers, Crowdfunding the PiDP-10 (and its size issues) — a thread on what a PDP-10 drew arch.

Chinook1994 · Chinook vs Marion Tinsley18 sockets

A single Silicon Graphics Challenge SMP server. The Chinook project page says 18 processors, 1 GB of RAM and 10 GB of disk; SGI's own 1994 press release says a Challenge XL with 16 processors, 1 GB of RAM and five 2 GB drives. The sources do not reconcile. A backup Onyx machine stood ready. P?

at the wall5,700 WPpoint

The rated maximum for the Challenge rack from SGI's owner's guide: 208-240 VAC, up to 30 A, 5,700 W, with one to three 1,900 W supplies. This is the chassis ceiling, not the draw of the configuration that played the match.

Sources: Silicon Graphics, Inc., POWER CHALLENGE and CHALLENGE XL Rackmount Owner's Guide (007-1735-050) arch.

at the wall3,200–5,700 WRrange

The lower bound is derived from the chassis heat dissipation given in the same guide, 10,900 BTU/hr, which works out to about 3.2 kW. The upper bound is the rated electrical maximum of 5,700 W. Nobody measured the actual draw of the 16-18 processor server, and room cooling is not included.

Sources: Silicon Graphics, Inc., POWER CHALLENGE and CHALLENGE XL Rackmount Owner's Guide (007-1735-050) arch.

Deep Blue1997 · Deep Blue versus Garry Kasparov510 nodes+chips

Thirty IBM RS/6000 SP nodes, one P2SC processor each, driving 480 custom chess chips — sixteen per node, on two Micro Channel cards of eight. The chip: 0.6-micron, three-metal, 5 V CMOS, about 1.5 million transistors, a 1.4 × 1.4 cm die, a 40–50 ns cycle. One chip searched 2–2.5 million positions per second; the system's design ceiling was roughly a billion positions per second and its sustained figure was 200 million. Node clock speed differs between sources: Campbell et al. say 120 MHz, BYTE of July 1997 says 135 MHz. M

chips only1,380 WRpoint

Reconstructed from the configuration; nobody put a meter on the wall. Two published numbers are added together: the chess chip's dissipation, given as 'about one watt' by the man who designed it (Hsu, IEEE Micro, 1999), times 480; plus 30 W per P2SC processor from the comparison table in BYTE of July 1997, times 30 nodes. That is 480 + 900 = 1,380 W. Silicon only: no memory, disks, SP switch, power-supply losses or cooling. The BYTE figure describes a 135 MHz P2SC, while the match machine ran at 120 MHz according to Campbell et al., so 900 W is closer to a ceiling than to a reading.

Sources: F.-h. Hsu, IBM's Deep Blue Chess Grandmaster Chips, 1999 arch. · Tom R. Halfhill, IBM's Deep Blue, 1997 arch. · M. Campbell, A. J. Hoane Jr., F.-h. Hsu, Deep Blue, 2002 arch.

at the wallno dataO?

There is no figure, and saying so is the honest answer. The '6.8 kW' that circulates online turns up in no source: not in Hsu (IEEE Micro, 1999), not in Campbell et al. (Artificial Intelligence, 2002), not in BYTE of 1997, not on IBM's own page, not in Wikipedia, not in the Chessprogramming Wiki, not in IEEE Spectrum's teardown of the chip. An exact-string search returns no origin, which is reason enough never to label it measured. The only figure with a traceable history is the 900 W in Jacques Mattheij's 2016 post, and that is 30 × 30 W from the BYTE table — the node processors and nothing else. What can be said: 1,380 W is a floor set by the silicon, and the installation at the wall drew appreciably more. How much more was never published.

Sources: F.-h. Hsu, IBM's Deep Blue Chess Grandmaster Chips, 1999 arch. · Tom R. Halfhill, IBM's Deep Blue, 1997 arch. · Jacques Mattheij, Another Way Of Looking At Lee Sedol vs AlphaGo, 2016 arch.

Logistello1997 · Logistello vs Takeshi Murakami1 chips

An ordinary single-processor PC. In his match report Buro calls it a "PentiumPro/233 (over-clocked)" system: Intel never shipped a 233 MHz Pentium Pro — the line topped out at 200 MHz — so this was overclocking. His overview of the program from the same year names the base machine as a PentiumPro/200 MHz running Linux with gcc 2.7.2.1, code in C. No special hardware of any kind: no multiprocessor node, no custom chips. P

chips only35 WPpoint

Intel's datasheet for the Pentium Pro at 200 MHz with 256 KB of L2 cache: thermal design power of 27.3 W typical, 35.0 W maximum. The match machine was overclocked to 233 MHz, a mode the datasheet does not cover, so the chip actually drew more.

Sources: Intel Corporation, Pentium Pro Processor at 150, 166, 180 and 200 MHz — datasheet, 1996 arch.

at the wall100–200 WOrange

Reconstructed from the class of machine: a mid-1990s desktop with one processor, one disk and a 200-250 W supply. No measurement of the match machine exists; the order of magnitude is a hundred watts and that is all this figure is good for.

Watson (DeepQA)2011 · Watson on Jeopardy!90 nodes

Ninety IBM Power 750 servers, each holding eight-core POWER7 processors at 3.55 GHz with four threads to a core — 2,880 cores and 16 TB of RAM in total. The whole corpus, some 200 million pages and about 4 TB, sat on disk and was held in memory: during play Watson had no connection to the internet. P

compute nodes175.5 kWRpoint

Reconstructed from the specification sheet; nobody metered the room. IBM publishes a maximum draw of 1,950 W per Power 750 Express server (2.0 kVA, 6,655 Btu/hr thermal output) in the operating environment table of Redpaper REDP-4638. Multiplied by 90 servers that is 175.5 kW. This is the nameplate ceiling for a fully loaded machine, not the average load during the game; the real draw was lower, and by how much IBM never said. Room cooling, UPS losses and external disk enclosures are not in the number.

Sources: IBM Redbooks, IBM Power 750 and 755 Technical Overview and Introduction (REDP-4638), table 1-1 “Operating environment”: 1,950 W maximum, 2.0 kVA, 6,655 BTU/hr, 2010 arch. · Wikipedia, IBM Watson, 2026 arch.

the facilityno dataO?

Not published.

Neither IBM nor any outside observer published a reading of the Watson installation's draw including cooling and UPS. The 80 kW and 200 kW figures that circulate do not lead back to a checkable source and are not repeated here.

Sources: IBM Redbooks, IBM Power 750 and 755 Technical Overview and Introduction (REDP-4638), table 1-1 “Operating environment”: 1,950 W maximum, 2.0 kVA, 6,655 BTU/hr, 2010 arch.

AlphaGo (the version DeepMind later named AlphaGo Lee)2016 · AlphaGo versus Lee Sedol48 chips

What stood behind the Seoul match was 48 first-generation tensor processing units spread across several machines. That is DeepMind's own statement, but it appeared only eighteen months after the match, in the AlphaGo Zero paper: 'AlphaGo Lee is the program that defeated Lee Sedol 4–1 in March 2016… This player was also distributed over many machines using 48 TPUs, rather than GPUs.' The much-quoted '1,920 CPUs and 280 GPUs' has nothing to do with this match: it is the largest row of the scalability table in the 2016 paper (Elo 3168), while the version that played Fan Hui in October 2015 was distributed over 1,202 CPUs and 176 GPUs. The confusion traces to DeepMind's pre-match line that AlphaGo would use 'about the same computing power' as against Fan Hui. How many machines and general-purpose processors served those 48 TPUs was never published. The hardware itself ran on Google Cloud in the United States; what sat in the Seoul room was Aja Huang, placing the stones. M

chips only1,920 WRpoint

Reconstructed from the configuration; nobody metered the match. 48 first-generation TPUs × 40 W. The forty watts is the measured busy power of a TPU die from Table 2 of Jouppi et al. (ISCA, 2017); the same table gives a 75 W nameplate TDP and 28 W at idle. That paper is directly on point: one of the convolutional networks in its benchmark set is AlphaGo's. Using TDP instead, the same multiplication gives 3,600 W. Accelerators only: the machines driving them are not counted, because their number was never made public.

Sources: N. P. Jouppi et al., In-Datacenter Performance Analysis of a Tensor Processing Unit (table 2: the TPU — 75 W TDP, 28 W idle, 40 W busy; a four-TPU server — 861 W TDP, 384 W busy), 2017 arch. · D. Silver, J. Schrittwieser, K. Simonyan, I. Antonoglou et al., Mastering the game of Go without human knowledge (Methods, “AlphaGo versions”: AlphaGo Lee — 48 TPUs), 2017 arch.

compute nodes4.6–10.3 kWOrange

Reconstructed from the configuration; nobody metered the match, and this is a floor. Jouppi et al. give a four-TPU server a measured busy draw of 384 W and a nameplate TDP of 861 W. Forty-eight TPUs make twelve such servers: from 12 × 384 ≈ 4.6 kW to 12 × 861 ≈ 10.3 kW. A range rather than a number, because the low end is a measurement under one particular load in Google's datacentre and the high end is a nameplate ceiling. Both are floors for the whole picture: DeepMind never published how many general-purpose machines served AlphaGo Lee's search, so none of their watts are in here.

Sources: N. P. Jouppi et al., In-Datacenter Performance Analysis of a Tensor Processing Unit (table 2: the TPU — 75 W TDP, 28 W idle, 40 W busy; a four-TPU server — 861 W TDP, 384 W busy), 2017 arch. · D. Silver, J. Schrittwieser, K. Simonyan, I. Antonoglou et al., Mastering the game of Go without human knowledge (Methods, “AlphaGo versions”: AlphaGo Lee — 48 TPUs), 2017 arch.

at the wallno dataO?

There is no figure. The famous contrast — 'AlphaGo a megawatt, Lee Sedol twenty watts' — goes back to a blog post Jacques Mattheij wrote during the match: he took the 1,920 CPUs and 280 GPUs from the scalability table of the 2016 paper, allowed 200 W apiece, arrived at roughly a megawatt and concluded 'fifty thousand times as much'. He gives no source for the 200 W, and more to the point that configuration never played Lee Sedol: Seoul was 48 TPUs. The megawatt is an estimate of the wrong machine. DeepMind has never published a measurement of its own.

Sources: Jacques Mattheij, Another Way Of Looking At Lee Sedol vs AlphaGo, 2016 arch. · D. Silver, A. Huang, C. J. Maddison, A. Guez, L. Sifre et al., Mastering the game of Go with deep neural networks and tree search, 2016 arch. · D. Silver, J. Schrittwieser, K. Simonyan, I. Antonoglou et al., Mastering the game of Go without human knowledge (Methods, “AlphaGo versions”: AlphaGo Lee — 48 TPUs), 2017 arch.

below this line: training · What it took to make the system able to play at all — before the match and outside it. It cannot be added to the match figure: training happened once, a game happens every time, and they are orders of magnitude apart.
trainingno dataO?

A separate quantity, and not a substitute for the match one. Training AlphaGo Lee took months on an incomparably larger fleet — the 2016 paper mentions 50 GPUs for network training — and the energy involved is orders of magnitude beyond five games. Neither DeepMind nor Google published a consumption figure for training, so this stays empty. Merging training with the match is the standard methodological error in this subject: different scope, different order of magnitude.

Sources: D. Silver, A. Huang, C. J. Maddison, A. Guez, L. Sifre et al., Mastering the game of Go with deep neural networks and tree search, 2016 arch. · D. Silver, J. Schrittwieser, K. Simonyan, I. Antonoglou et al., Mastering the game of Go without human knowledge (Methods, “AlphaGo versions”: AlphaGo Lee — 48 TPUs), 2017 arch.

Libratus2017 · Libratus versus four professionals600 nodes

Regular-memory nodes of the Bridges supercomputer: two fourteen-core Intel Xeon E5-2695 v3 processors at 2.3 GHz and 128 GB of RAM apiece, so 28 cores to a node. How many nodes depends on which part of the system is meant, and the sources differ because they describe different things. Brown and Sandholm (IJCAI-18): equilibrium finding and the self-improvement module ran on 196 nodes, with the agent using only 14 of each node's 28 cores; the subgame solver took 50 nodes per game in progress. Carnegie Mellon's press release of 31 January 2017 says roughly 600 of Bridges' 846 nodes were in use during the competition. These do not contradict each other: 600 is the whole occupied machine across a playing day, parallel tables and the overnight recomputation included, while 50 and 196 are single modules. The record carries 600 — that is what was running while the men sat at the table. R

at the wallno dataO?

Not published.

Neither the Pittsburgh Supercomputing Center nor the Libratus authors published what the occupied share of Bridges drew at the wall, either per playing day or across the match. All that exists is a reconstruction from the processors' nameplate TDP, and that sets a floor, not a ceiling.

Sources: N. Brown, T. Sandholm, Libratus: The Superhuman AI for No-Limit Poker (Demonstration) (the “Agent Construction” section: about 25 million core-hours in all, of which some 13 million went on search experiments, 6 million on abstraction and equilibrium finding, 3 million on subgame solving and 3 million on self-improvement; 196 Bridges nodes of 28 cores and 128 GB, the agent itself holding 14 cores; the subgame solver ran on 50 nodes per game), 2018 arch.

the match12–144 kWRrange

Reconstructed from the configuration; nobody put a meter on the wall. The Xeon E5-2695 v3 carries a 120 W nameplate TDP (Intel specification) and a Bridges node holds two of them, so 240 W of silicon per node. The low end of the range is the 50 nodes per game in progress given by Brown and Sandholm: 50 × 240 = 12 kW. The high end is the roughly 600 occupied nodes in Carnegie Mellon's press release: 600 × 240 = 144 kW. Processors only — memory, disks, the Omni-Path fabric, power-supply losses and room cooling are all outside the number, so the draw at the wall was appreciably larger. Total match energy converts badly from core hours: Brown and Sandholm allot about 3 million core hours to the subgame solver and the same again to self-improvement, but never say what share of those hours belonged to the 120,000 hands rather than to debugging. Read the core hour two ways — 14 used cores per node, or all 28 — and 3 million core hours come to between 26 and 51 MWh, which is 0.8 to 1.5 MJ per hand. That is one and a half orders of magnitude away from Pluribus's 4.8 kJ per hand, and it is why these two records belong side by side.

Sources: N. Brown, T. Sandholm, Libratus: The Superhuman AI for No-Limit Poker (Demonstration) (the “Agent Construction” section: about 25 million core-hours in all, of which some 13 million went on search experiments, 6 million on abstraction and equilibrium finding, 3 million on subgame solving and 3 million on self-improvement; 196 Bridges nodes of 28 cores and 128 GB, the agent itself holding 14 cores; the subgame solver ran on 50 nodes per game), 2018 arch. · Byron Spice, Garrett Allen, Carnegie Mellon Artificial Intelligence Beats Top Poker Pros (120,000 hands, a margin of $1,766,250 in chips, a $200,000 prize pool, “approximately 600 of Bridges' 846 compute nodes”, the strategy recomputed overnight), 2017 arch. · Intel, Intel Xeon Processor E5-2695 v3 (35M Cache, 2.30 GHz) — Product Specifications: 14 cores, 28 threads, 2.30 GHz base, 3.30 GHz turbo, 22 nm, 120 W TDP, launched Q3 2014, 2014 arch.

below this line: training · What it took to make the system able to play at all — before the match and outside it. It cannot be added to the match figure: training happened once, a game happens every time, and they are orders of magnitude apart.
training47 kWRpoint

Reconstructed from the configuration. 196 Bridges nodes × 240 W of silicon = 47.04 kW is what the processors pulled while the strategy's blueprint was being computed. Brown and Sandholm allot about 6 million core hours to abstraction and equilibrium finding. Counting a core hour against the 14 cores the agent actually occupied, that is 429,000 node-hours — some 91 days of all 196 nodes running flat out, and roughly 103 MWh. Counting against all 28 cores of a node halves it to about 51 MWh. The paper does not say how the core hours were tallied, so the honest span is 51–103 MWh, silicon only. For scale: the whole Libratus effort came to about 25 million core hours by the same paper's reckoning, 13 million of which went on exploratory experiments and evaluation — 214–429 MWh by the same method.

Sources: N. Brown, T. Sandholm, Libratus: The Superhuman AI for No-Limit Poker (Demonstration) (the “Agent Construction” section: about 25 million core-hours in all, of which some 13 million went on search experiments, 6 million on abstraction and equilibrium finding, 3 million on subgame solving and 3 million on self-improvement; 196 Bridges nodes of 28 cores and 128 GB, the agent itself holding 14 cores; the subgame solver ran on 50 nodes per game), 2018 arch. · Pittsburgh Supercomputing Center, Bridges RSM (Regular Shared Memory) nodes: two 14-core Intel Xeon E5-2695 v3 at 2.3 GHz and 128 GB of memory per node, 2015 arch. · Intel, Intel Xeon Processor E5-2695 v3 (35M Cache, 2.30 GHz) — Product Specifications: 14 cores, 28 threads, 2.30 GHz base, 3.30 GHz turbo, 22 nm, 120 W TDP, launched Q3 2014, 2014 arch.

Pluribus2019 · Pluribus at a six-handed table2 sockets

In play Pluribus ran on two Intel Haswell E5-2695 v3 processors and used under 128 GB of memory — one server of the same kind Bridges is built from, one rather than six hundred. No accelerators, no cluster. The blueprint was computed separately and in advance: eight days on a 64-core server, 12,400 core hours in total, under 512 GB of memory; the paper does not name that training server's processor. The authors sized the abstraction deliberately so that the playing machine would fit in 128 GB with a compressed blueprint held in memory. The comparison they draw themselves: AlphaGo used 1,920 CPUs and 280 GPUs, Deep Blue 480 custom chips, Libratus 100 CPUs. M

the match240 WRpoint4,800 J per hand

Reconstructed from the configuration; nobody metered the table. Two Xeon E5-2695 v3 processors × 120 W nameplate TDP = 240 W. Joules per hand: 240 W × 20 s = 4,800 J. The twenty seconds is Pluribus's average pace against copies of itself, as stated in the paper; at a table with people the machine waited for people, and how much of that time it spent computing was never published. So the figure describes work done, not wall-clock time at the table. A single decision — one subgame search — took between 1 and 33 seconds, which is 0.24 to 7.9 kJ. All 10,000 hands of the five-professionals experiment at that pace come to about 48 MJ, or 13 kWh: roughly what a domestic refrigerator uses in a day. Processors only; memory, disk and power-supply losses are not in the number.

Sources: N. Brown, T. Sandholm, Superhuman AI for multiplayer poker (p. 887: the blueprint strategy — 8 days on a 64-core server, 12,400 core-hours, under 512 GB of memory, “about $144 to produce” at spot cloud prices; play ran on two Intel Haswell E5-2695 v3 and under 128 GB; subgame search took 1 to 33 s, averaging 20 s per hand against copies of itself), 2019 arch. · Intel, Intel Xeon Processor E5-2695 v3 (35M Cache, 2.30 GHz) — Product Specifications: 14 cores, 28 threads, 2.30 GHz base, 3.30 GHz turbo, 22 nm, 120 W TDP, launched Q3 2014, 2014 arch.

below this line: training · What it took to make the system able to play at all — before the match and outside it. It cannot be added to the match figure: training happened once, a game happens every time, and they are orders of magnitude apart.
trainingno dataO?

There are no watts here and nothing honest to put in their place: the paper says '64-core server' without naming the processor, and without the model there is no nameplate TDP to use. What is firmly published: eight days of computing, 12,400 core hours, under 512 GB of memory. The comparison worth making is in core hours rather than watts, and there the contrast needs no reconstruction at all: Libratus spent about 6 million core hours on its blueprint alone and about 25 million on the whole effort — 484 and 2,016 times more. Pluribus's training fits inside one server and one working week, and that fact, rather than the result at the table, is what makes this record worth keeping.

Sources: N. Brown, T. Sandholm, Superhuman AI for multiplayer poker (p. 887: the blueprint strategy — 8 days on a 64-core server, 12,400 core-hours, under 512 GB of memory, “about $144 to produce” at spot cloud prices; play ran on two Intel Haswell E5-2695 v3 and under 128 GB; subgame search took 1 to 33 s, averaging 20 s per hand against copies of itself), 2019 arch. · N. Brown, T. Sandholm, Libratus: The Superhuman AI for No-Limit Poker (Demonstration) (the “Agent Construction” section: about 25 million core-hours in all, of which some 13 million went on search experiments, 6 million on abstraction and equilibrium finding, 3 million on subgame solving and 3 million on self-improvement; 196 Bridges nodes of 28 cores and 128 GB, the agent itself holding 14 cores; the subgame solver ran on 50 nodes per game), 2018 arch.

KataGo and the third-party build JBXKata005; separately, Leela Zero2023 · Kellin Pelrine versus KataGo

Nobody published the machine side's hardware: the bots lived on the KGS server, which belonged to somebody else and was outside the authors' control. What is on record is not nodes but search settings, and those matter more: KataGo and Leela Zero played at 100,000 visits per move — full search, not a bare policy network. This is the load-bearing detail. Without search, the adversarial policy beats KataGo in over 99% of games; with search the job is markedly harder, which is precisely why the human's games count for something. The opponent in the fifteen-game series was JBXKata005, an independent KataGo build unaffiliated with the authors and the strongest bot then playable on KGS; its settings are not given in the paper. The KataGo network used for the main machine experiments is named exactly — b40c256-s11840935168-d2898845681 — but no network number is given for the games against the human. This is where the imprecision that travelled through the press comes from: the Financial Times of 19 February 2023 ran the series as 'man beats machine at Go', when the fifteen games were played against a third-party build. M

the matchno dataO?

Not published.

The games ran on the public KGS server, on other people's machines. What hardware turned JBXKata005, KataGo and Leela Zero is stated neither in the paper nor on the project site. The only measurable quantity on the machine side is the 100,000 visits per move used by KataGo and Leela Zero, and that describes work, not watts. Filling the gap with the TDP of some plausible graphics card would be invention.

Sources: T. T. Wang, A. Gleave, T. Tseng, K. Pelrine, N. Belrose, J. Miller, M. D. Dennis, Y. Duan, V. Pogrebniak, S. Levine, S. Russell, Adversarial Policies Beat Superhuman Go AIs (appendix G.2 — the attack carried over to a human: 14 games of 15 against JBXKata005, wins over KataGo and Leela Zero at 100,000 visits, handicaps of 3, 5 and 9 stones, “without any algorithmic assistance”; appendix D — 2,223.2 V100-days for the adversarial policy, 1.01 million self-play games, no fewer than 15,881 V100-days to train KataGo, 6,730 of them the “third major run”; section 5.4 — KataGo fine-tuned against the attack from mid-December 2022), 2023 arch. · FAR AI, Adversarial Policies in Go — the game records, the human-played ones among them, 2023 arch.

below this line: training · What it took to make the system able to play at all — before the match and outside it. It cannot be added to the match figure: training happened once, a game happens every time, and they are orders of magnitude apart.
training8.4–13.8 kWRrange

This is KataGo's own training, not the games. Reconstructed from the configuration: the KataGo repository documents the 'third major run' as 14 days on 28 V100 accelerators, 24 days on 36, and 119 days on 46 — quoted in the adversarial-attack paper, which totals it at 6,730 V100 GPU-days. The watt range is those same three phases: 28 × 300 = 8.4 kW at the start and 46 × 300 = 13.8 kW at the end, using the V100's nameplate TDP in SXM2 form (the PCIe card is 250 W). Those 6,730 days come to 40–48 MWh, accelerator silicon only — nothing for the machines holding them and nothing for cooling. The network actually attacked was trained longer: the authors give a lower bound of 15,881 V100 GPU-days, so not less than 114 MWh by the same arithmetic — but that is a conversion unit rather than real V100s, since the later training was distributed across volunteers' assorted hardware.

Sources: T. T. Wang, A. Gleave, T. Tseng, K. Pelrine, N. Belrose, J. Miller, M. D. Dennis, Y. Duan, V. Pogrebniak, S. Levine, S. Russell, Adversarial Policies Beat Superhuman Go AIs (appendix G.2 — the attack carried over to a human: 14 games of 15 against JBXKata005, wins over KataGo and Leela Zero at 100,000 visits, handicaps of 3, 5 and 9 stones, “without any algorithmic assistance”; appendix D — 2,223.2 V100-days for the adversarial policy, 1.01 million self-play games, no fewer than 15,881 V100-days to train KataGo, 6,730 of them the “third major run”; section 5.4 — KataGo fine-tuned against the attack from mid-December 2022), 2023 arch. · NVIDIA, NVIDIA Tesla V100 GPU Accelerator — Datasheet: 300 W TDP in the SXM2 form factor, 250 W in the PCIe one, 2018 arch.

trainingno dataO

This is the training of the attacking policy — the price of finding the very strategy the human later won with. The authors give the hardware actually used: 61.8 A4000 GPU-days, 348.6 A6000, 299.7 A100 40GB and 222.9 A100 80GB, which they themselves convert to 2,223.2 V100 GPU-days — about 14% of the compute that went into training KataGo, and roughly 1.01 million self-play games. Applying those accelerators' nameplate TDPs (140 W for the A4000, 300 for the A6000, 250–400 for the A100 40GB, 300–500 for the A100 80GB, the spread coming from PCIe and SXM variants the paper does not distinguish) gives something on the order of 6–8 MWh of silicon. This is an order of magnitude, not a measurement: how many accelerators ran at once is not stated, so there is no watt figure here, only a total.

Sources: T. T. Wang, A. Gleave, T. Tseng, K. Pelrine, N. Belrose, J. Miller, M. D. Dennis, Y. Duan, V. Pogrebniak, S. Levine, S. Russell, Adversarial Policies Beat Superhuman Go AIs (appendix G.2 — the attack carried over to a human: 14 games of 15 against JBXKata005, wins over KataGo and Leela Zero at 100,000 visits, handicaps of 3, 5 and 9 stones, “without any algorithmic assistance”; appendix D — 2,223.2 V100-days for the adversarial policy, 1.01 million self-play games, no fewer than 15,881 V100-days to train KataGo, 6,730 of them the “third major run”; section 5.4 — KataGo fine-tuned against the attack from mid-December 2022), 2023 arch. · NVIDIA, NVIDIA Tesla V100 GPU Accelerator — Datasheet: 300 W TDP in the SXM2 form factor, 250 W in the PCIe one, 2018 arch.

What we do not know
An absence as a result of the research, not a hole in the layout

Of the 22 quantities, 8 have no number. These are not cells somebody forgot: each was looked for, and what follows says where the search went and what it failed to find. The pattern is telling — the blanks sit where the boundary is widest: at the wall, at the facility, and in training. Those are the three figures most often quoted.

at the wallno dataO?Deep Blue · 1997

There is no figure, and saying so is the honest answer. The '6.8 kW' that circulates online turns up in no source: not in Hsu (IEEE Micro, 1999), not in Campbell et al. (Artificial Intelligence, 2002), not in BYTE of 1997, not on IBM's own page, not in Wikipedia, not in the Chessprogramming Wiki, not in IEEE Spectrum's teardown of the chip. An exact-string search returns no origin, which is reason enough never to label it measured. The only figure with a traceable history is the 900 W in Jacques Mattheij's 2016 post, and that is 30 × 30 W from the BYTE table — the node processors and nothing else. What can be said: 1,380 W is a floor set by the silicon, and the installation at the wall drew appreciably more. How much more was never published.

Sources: F.-h. Hsu, IBM's Deep Blue Chess Grandmaster Chips, 1999 arch. · Tom R. Halfhill, IBM's Deep Blue, 1997 arch. · Jacques Mattheij, Another Way Of Looking At Lee Sedol vs AlphaGo, 2016 arch.

the facilityno dataO?Watson · 2011

Not published.

Neither IBM nor any outside observer published a reading of the Watson installation's draw including cooling and UPS. The 80 kW and 200 kW figures that circulate do not lead back to a checkable source and are not repeated here.

Sources: IBM Redbooks, IBM Power 750 and 755 Technical Overview and Introduction (REDP-4638), table 1-1 “Operating environment”: 1,950 W maximum, 2.0 kVA, 6,655 BTU/hr, 2010 arch.

at the wallno dataO?AlphaGo · 2016

There is no figure. The famous contrast — 'AlphaGo a megawatt, Lee Sedol twenty watts' — goes back to a blog post Jacques Mattheij wrote during the match: he took the 1,920 CPUs and 280 GPUs from the scalability table of the 2016 paper, allowed 200 W apiece, arrived at roughly a megawatt and concluded 'fifty thousand times as much'. He gives no source for the 200 W, and more to the point that configuration never played Lee Sedol: Seoul was 48 TPUs. The megawatt is an estimate of the wrong machine. DeepMind has never published a measurement of its own.

Sources: Jacques Mattheij, Another Way Of Looking At Lee Sedol vs AlphaGo, 2016 arch. · D. Silver, A. Huang, C. J. Maddison, A. Guez, L. Sifre et al., Mastering the game of Go with deep neural networks and tree search, 2016 arch. · D. Silver, J. Schrittwieser, K. Simonyan, I. Antonoglou et al., Mastering the game of Go without human knowledge (Methods, “AlphaGo versions”: AlphaGo Lee — 48 TPUs), 2017 arch.

trainingno dataO?AlphaGo · 2016

A separate quantity, and not a substitute for the match one. Training AlphaGo Lee took months on an incomparably larger fleet — the 2016 paper mentions 50 GPUs for network training — and the energy involved is orders of magnitude beyond five games. Neither DeepMind nor Google published a consumption figure for training, so this stays empty. Merging training with the match is the standard methodological error in this subject: different scope, different order of magnitude.

Sources: D. Silver, A. Huang, C. J. Maddison, A. Guez, L. Sifre et al., Mastering the game of Go with deep neural networks and tree search, 2016 arch. · D. Silver, J. Schrittwieser, K. Simonyan, I. Antonoglou et al., Mastering the game of Go without human knowledge (Methods, “AlphaGo versions”: AlphaGo Lee — 48 TPUs), 2017 arch.

at the wallno dataO?Libratus · 2017

Not published.

Neither the Pittsburgh Supercomputing Center nor the Libratus authors published what the occupied share of Bridges drew at the wall, either per playing day or across the match. All that exists is a reconstruction from the processors' nameplate TDP, and that sets a floor, not a ceiling.

Sources: N. Brown, T. Sandholm, Libratus: The Superhuman AI for No-Limit Poker (Demonstration) (the “Agent Construction” section: about 25 million core-hours in all, of which some 13 million went on search experiments, 6 million on abstraction and equilibrium finding, 3 million on subgame solving and 3 million on self-improvement; 196 Bridges nodes of 28 cores and 128 GB, the agent itself holding 14 cores; the subgame solver ran on 50 nodes per game), 2018 arch.

trainingno dataO?Pluribus · 2019

There are no watts here and nothing honest to put in their place: the paper says '64-core server' without naming the processor, and without the model there is no nameplate TDP to use. What is firmly published: eight days of computing, 12,400 core hours, under 512 GB of memory. The comparison worth making is in core hours rather than watts, and there the contrast needs no reconstruction at all: Libratus spent about 6 million core hours on its blueprint alone and about 25 million on the whole effort — 484 and 2,016 times more. Pluribus's training fits inside one server and one working week, and that fact, rather than the result at the table, is what makes this record worth keeping.

Sources: N. Brown, T. Sandholm, Superhuman AI for multiplayer poker (p. 887: the blueprint strategy — 8 days on a 64-core server, 12,400 core-hours, under 512 GB of memory, “about $144 to produce” at spot cloud prices; play ran on two Intel Haswell E5-2695 v3 and under 128 GB; subgame search took 1 to 33 s, averaging 20 s per hand against copies of itself), 2019 arch. · N. Brown, T. Sandholm, Libratus: The Superhuman AI for No-Limit Poker (Demonstration) (the “Agent Construction” section: about 25 million core-hours in all, of which some 13 million went on search experiments, 6 million on abstraction and equilibrium finding, 3 million on subgame solving and 3 million on self-improvement; 196 Bridges nodes of 28 cores and 128 GB, the agent itself holding 14 cores; the subgame solver ran on 50 nodes per game), 2018 arch.

trainingno dataOKataGo · 2023

This is the training of the attacking policy — the price of finding the very strategy the human later won with. The authors give the hardware actually used: 61.8 A4000 GPU-days, 348.6 A6000, 299.7 A100 40GB and 222.9 A100 80GB, which they themselves convert to 2,223.2 V100 GPU-days — about 14% of the compute that went into training KataGo, and roughly 1.01 million self-play games. Applying those accelerators' nameplate TDPs (140 W for the A4000, 300 for the A6000, 250–400 for the A100 40GB, 300–500 for the A100 80GB, the spread coming from PCIe and SXM variants the paper does not distinguish) gives something on the order of 6–8 MWh of silicon. This is an order of magnitude, not a measurement: how many accelerators ran at once is not stated, so there is no watt figure here, only a total.

Sources: T. T. Wang, A. Gleave, T. Tseng, K. Pelrine, N. Belrose, J. Miller, M. D. Dennis, Y. Duan, V. Pogrebniak, S. Levine, S. Russell, Adversarial Policies Beat Superhuman Go AIs (appendix G.2 — the attack carried over to a human: 14 games of 15 against JBXKata005, wins over KataGo and Leela Zero at 100,000 visits, handicaps of 3, 5 and 9 stones, “without any algorithmic assistance”; appendix D — 2,223.2 V100-days for the adversarial policy, 1.01 million self-play games, no fewer than 15,881 V100-days to train KataGo, 6,730 of them the “third major run”; section 5.4 — KataGo fine-tuned against the attack from mid-December 2022), 2023 arch. · NVIDIA, NVIDIA Tesla V100 GPU Accelerator — Datasheet: 300 W TDP in the SXM2 form factor, 250 W in the PCIe one, 2018 arch.

the humanno dataO?

The increment from hard mental work. There is no reliable figure, and that is itself the finding: the increment is small enough to be hard to measure at all.

The most direct test was done seventy years ago and has not been overturned. Sokoloff, Mangold, Wechsler, Kennedy and Kety (Journal of Clinical Investigation, 1955) measured cerebral blood flow and oxygen consumption in 13 people at rest and while doing mental arithmetic. The EEG changed unmistakably during the arithmetic — a shift to low-voltage, high-frequency activity — and pulse and blood pressure rose, while the brain's oxygen consumption did not move: in their words, the control and arithmetic values were identical. The authors state the conclusion worth remembering the paper for: neither the degree of mental activity nor the EEG pattern correlates with total cerebral oxygen consumption.

Modern imaging did not overturn this so much as sharpen it. Locally, blood flow in a working region rises a great deal; across the whole brain, the task-evoked increase is usually under 5% of the ongoing baseline (Raichle and Mintun, Annual Review of Neuroscience, 2006). Five per cent of 20 W is under a watt, and that is a ceiling rather than a reading.

What is not here and will not be: the claim that a grandmaster burns 6,000 kilocalories a day. It comes from an ESPN piece of 2019 quoting Robert Sapolsky, and traces back to a tripled breathing rate multiplied by 2,000 kcal. No calorimeter appears anywhere in that chain. What does rise during a hard game is sympathetic arousal — heart rate, blood pressure, breathing, muscle tone. That raises the whole body's expenditure, not the brain's, and how much it raises it over a chess match has never been published.

Sources: L. Sokoloff, R. Mangold, R. L. Wechsler, C. Kennedy, S. S. Kety, The Effect of Mental Arithmetic on Cerebral Circulation and Metabolism (13 subjects; doing arithmetic in the head shifts the EEG and raises pulse and blood pressure, while cerebral blood flow and the brain's oxygen consumption do not move; control CMRO₂ 3.3–4.0 ml O₂ per 100 g per minute), 1955 arch. · M. E. Raichle, M. A. Mintun, Brain Work and Brain Imaging (the increase in the brain's energy use that a task brings about is small next to the baseline — usually under 5%), 2006 arch. · Aishwarya Kumar, The grandmaster diet: How to lose weight while barely moving (where the “6,000 calories a day” figure comes from, credited to Robert Sapolsky; no calorimetry underneath it), 2019 arch. · Andrew Gelman, About that bogus claim that “chess grandmasters” burn 6000 calories per day (where the number came from: respiration measured on 11 ordinary players, the highest reading passed off as the average, multiplied against 2,000 kcal), 2025 arch.

What follows from it

The thermodynamic mismatch has not gone anywhere: even the silicon-only floor for the 1997 machine is an order of magnitude above a seated person's entire heat output, and Watson's is three orders above it. But its size depends on what is being set against what, and the spread of defensible answers is wider than the effect itself in half these cases. A comparison that does not name its scope reports nothing except where the person making it stands.

Nor is the machine's curve single or monotonic. From 1979 to 2011 it rose — a cabinet, then ninety servers — and between 2017 and 2019, in poker, it fell from a hundred kilowatts to two hundred and forty watts for the same result at the table. The human line has not moved a watt in forty-four years: a hundred and something, on the same thermal comfort tables as in 1979.

The data is the site's system records, assembled on 12 August 2026 and checked by a script of their own. The neighbouring axes of the same dataset: The Price of a Move for money, the processor catalogue for the hardware.