The Long Road - A History of Computing Machines

14.0 What this chapter gives you#

1. You will be able to tell the story of computing from tally sticks to the machines of 2026, in order, with correct years and names.
2. You will be able to explain what a punched card did in a loom in 1804, and why that counts as the first machine following stored instructions.
3. You will be able to say what Charles Babbage designed, what Ada Lovelace wrote in 1843, and where the honest scholarly argument about her sits.
4. You will be able to explain a Turing machine, the halting problem and the phrase “Turing complete” to somebody who knows no mathematics.
5. You will be able to name the wartime machines of 1941 to 1945 and say exactly what each one could and could not do.
6. You will be able to explain why the stored-program idea of 1945 changed everything, and who deserves credit for it.
7. You will be able to trace the chain: transistor, integrated circuit, microprocessor, personal computer, graphical screen, network, phone.
8. You will be able to say why the industry stopped raising clock speed around 2005 and started adding cores instead.
9. You will be able to correct the five most common history mistakes people repeat about computers.
10. You will have one master timeline you can read in five minutes.

14.1 Counting before machines#

PLAIN14.1.1 in simple words#

1. For most of human history, counting aids were sticks, stones and beads.
2. A tally stick is a stick with notches cut into it. One notch, one thing.
3. Some tally sticks were split lengthwise so two people each kept half. The matching notches proved the debt. England used such sticks for tax records for centuries, right up to 1826.
4. An abacus is a frame of beads on rods. Each rod is a decimal place.
5. The beads do not calculate. The person calculates. The beads hold the partial answer so the person does not have to remember it.
6. That is the first big idea in this whole chapter: a machine that remembers for you is already useful, even if it does no thinking.
7. In 1901 sponge divers found a corroded lump of bronze in a shipwreck off the Greek island of Antikythera.
8. Inside were at least 30 meshing gears. It is called the Antikythera mechanism, and it modelled the sky.
9. Turn a handle and pointers showed the positions of the Sun and Moon, the phase of the Moon, and when eclipses were due.
10. Nothing of that mechanical complexity is known again for over a thousand years. It was a dead end, not a starting point.

PLAIN14.1.2 a picture in your head#

1. Think of a kitchen with no notepad.
2. You are adding a long shopping list in your head. By item nine you have lost item three.
3. Now somebody hands you a row of ten cups and a bag of stones.
4. You drop stones into cups to hold each part of the sum. Your head is free.
5. The cups do not add. You add. The cups only refuse to forget.
6. Napier’s bones, the slide rule and the abacus are all cups and stones. They store a partial result, or they turn a hard job into an easy one.
7. The Antikythera mechanism is different. It is a clockwork sky. The gear ratios are the astronomy, baked into metal once and for all.

Where this comparison breaks: cups and stones can hold any number you like, but a gear train can only ever do the one calculation its teeth were cut for. To change the question you must cut new gears. That difference, fixed function versus changeable instructions, is the whole story of the next 2,000 years.

PLAIN14.1.3 a worked example#

1. John Napier published Napier’s bones in 1617, in a book called Rabdologiae, printed in Edinburgh. He died that same year.
2. Each bone is a rod carrying one column of the multiplication table.
3. To multiply 4,875 by 7, lay out the bones for 4, 8, 7 and 5, and read row 7.
4. Row 7 gives, in each bone, the digits of 7 times that digit, split by a diagonal line into tens and units.

Bones for   4     8     7     5
Row 7:    2/8   5/6   4/9   3/5

Add along the diagonals, right to left:
  units:            5
  next:       3 + 9 = 12  -> write 2, carry 1
  next:   1 + 4 + 6 = 11  -> write 1, carry 1
  next:   1 + 5 + 8 = 14  -> write 4, carry 1
  next:       1 + 2 = 3
Result: 34125

5. Check: 4,875 times 7 is 34,125. The bones turned four multiplications into four small additions.
6. That is the trick every pre-electronic aid uses. Convert a hard operation into an easier one you can already do.

PLAIN14.1.4 what is really happening inside#

1. Napier’s other invention, logarithms, published in 1614, does the same trick at a deeper level.
2. A logarithm turns multiplication into addition. Multiply two numbers by adding their logarithms and looking up the answer.
3. Edmund Gunter engraved a logarithmic scale on a ruler around 1620. William Oughtred put two such scales side by side around 1622 and slid one along the other. That is the slide rule.
4. Sliding one scale by a distance is adding the logarithms. The mechanism does the addition by geometry.
5. In 1623 Wilhelm Schickard, a professor at Tübingen, described a “calculating clock” in letters to the astronomer Johannes Kepler.
6. It used Napier’s bones mounted on cylinders for multiplication and a geared adder underneath. The machine being built for Kepler burned in a fire before delivery. Schickard died of plague in 1635.
7. In 1642 the nineteen-year-old Blaise Pascal built the Pascaline to help his father, a tax official, add long columns of money.
8. It added and subtracted with toothed wheels, and it carried automatically: when a wheel passed nine, a weighted lever tipped and nudged the next wheel.
9. Around fifty were made. About nine survive.
10. Gottfried Wilhelm Leibniz went further. He showed a wooden model of his stepped reckoner to the Royal Society in London on 1 February 1673.
11. It was the first design that could do all four operations: add, subtract, multiply and divide.
12. In 1703 Leibniz published a short paper, Explication de l’Arithmétique Binaire, setting out arithmetic using only 0 and 1.
13. He did not build a binary machine. But every machine in the rest of this book runs on the notation he wrote down in 1703.

TECHNICAL14.1.5 the engineer’s version#

1. The Antikythera mechanism carries at least 30 surviving bronze gears; the largest has 223 teeth, encoding the 223-month Saros eclipse cycle.
2. Dating is contested. Published estimates cluster around 150 to 100 BCE, with arguments for 205 BCE and for about 87 BCE. The shipwreck itself dates to roughly 70 to 60 BCE. Experts disagree; treat “around 100 BCE” as a reasonable midpoint, not a settled fact.
3. Its lunar train uses a pin-and-slot on two gears of 53 teeth to produce a varying angular rate, modelling the Moon’s elliptical motion. That is a mechanical implementation of a first-order anomaly correction.
4. The Pascaline’s carry mechanism, the sautoir, is gravity-driven. Its drawback is that a long carry chain, such as 999,999 plus 1, must lift several weights at once, so torque rises with carry length.
5. Leibniz’s key part is the stepped drum (Staffelwalze): a cylinder with nine teeth of increasing length. Shifting a gear along the drum selects how many teeth engage per turn, which multiplies by a chosen digit.
6. The stepped drum stayed in production hardware until the twentieth century, in the Thomas Arithmometer from 1820 and later in the Curta of 1948.

7. Precision limits, not ideas, blocked all of these. Leibniz’s machine had a carry fault that was only fully understood when it was studied in the twentieth century.
8. Standard versus convention: none of this is standardized. These are one-off artefacts. The first true standard in computing history is the punched card format, and that comes in section 14.4.

WORDS14.1.6 remember these#

1. Tally stick — a notched stick counting things — a unary record medium.
2. Abacus — beads on rods — a positional manual accumulator.
3. Antikythera mechanism — an ancient geared sky model — a bronze analogue astronomical calculator, roughly 100 BCE.
4. Napier’s bones — rods holding times tables — lattice multiplication aid,
5. Logarithm — a number that turns times into plus — the exponent to which a fixed base must be raised.
6. Slide rule — two sliding log scales — an analogue multiplier, about 1622.
7. Pascaline — Pascal’s adding box of 1642 — geared decimal adder with gravity carry.
8. Stepped reckoner — Leibniz’s four-function machine, 1673 — stepped-drum mechanical calculator.
9. Binary — counting with only 0 and 1 — base-2 positional notation, set out by Leibniz in 1703.

14.2 Machines that follow instructions#

PLAIN14.2.1 in simple words#

1. A calculator does one job. To change the job you change the machine.
2. The next idea is bigger: leave the machine alone, and change a stack of cards instead.
3. In 1804 Joseph Marie Jacquard, in Lyon in France, patented an attachment that sat on top of a weaving loom.
4. Weaving a picture into cloth means lifting a different set of threads for every single row of the picture.
5. Before Jacquard, a boy sat inside the loom pulling cords by hand, one row at a time, all day, for weeks.
6. Jacquard replaced the boy with a chain of stiff cards, each punched with holes.
7. Rods press against a card. Where there is a hole, a rod goes through and its thread lifts. Where there is card, the rod is blocked and the thread stays.
8. One card equals one row of the picture. The chain of cards equals the whole picture.
9. To weave a different picture you do not rebuild the loom. You hang a different chain of cards.
10. That separation, machine on one side and instructions on the other, is the single most important idea in this book.

PLAIN14.2.2 a picture in your head#

1. Think of a music box with a spiked cylinder.
2. The spikes pluck the metal teeth. The cylinder is the tune.
3. A cheap music box has the cylinder fixed. It plays one tune for ever.
4. A better music box lets you pull the cylinder out and push a new one in. Same box, new tune.
5. Jacquard’s loom is that better music box, except the tune is a picture and the cylinder is a loop of punched cards.
6. A player piano works exactly this way too, with a punched paper roll.

Where this comparison breaks: a music box or piano roll only ever plays forwards. It cannot skip back, repeat a bar four times, or choose a different bar depending on what happened. The Jacquard chain has the same limit. Real programs need to jump, loop and decide. That missing power is exactly what Babbage tried to add in 1837, and what nobody achieved in working hardware until the 1940s.

PLAIN14.2.3 a worked example#

1. Take a tiny loom with 8 threads, weaving a picture 5 rows tall.
2. Punch 1 means “lift this thread”. No punch, shown as 0, means “leave it”.

Card 1:  0 0 1 1 1 1 0 0
Card 2:  0 1 1 0 0 1 1 0
Card 3:  1 1 0 0 0 0 1 1
Card 4:  0 1 1 0 0 1 1 0
Card 5:  0 0 1 1 1 1 0 0

3. Read the 1s down the page and a diamond shape appears in the cloth.
4. Five cards, eight holes each: 40 bits of information decide a picture.
5. Real Jacquard cards of the period commonly carried several hundred hole positions per card, and a single elaborate portrait could need tens of thousands of cards.
6. The famous woven portrait of Jacquard himself, made in the 1830s, used around 24,000 cards for one image.
7. Charles Babbage owned a copy of that portrait and showed it to visitors. That is a direct, documented link from weaving to computing.

PLAIN14.2.4 what is really happening inside#

1. Jacquard did not invent this from nothing. He combined three earlier French ideas.
2. Basile Bouchon used a punched paper tape for loom control in 1725.
3. Jean-Baptiste Falcon replaced the fragile tape with a chain of stiff cards in 1728.
4. Jacques de Vaucanson built an automatic loom around 1740 using a perforated cylinder.
5. Jacquard’s 1804 machine put the strong parts together and made it reliable and cheap enough for real workshops.
6. Mechanically: a rectangular card presses against a bank of spring-loaded needles. Each needle controls a hook. Each hook lifts one warp thread.
7. Hole present: needle passes through, hook stays engaged with the lifting knife, thread rises.
8. No hole: needle is pushed back, hook is displaced sideways, the knife misses it, thread stays down.
9. The card is therefore a row of yes/no decisions, read in parallel, once per row of cloth. It is a memory device made of cardboard.
10. Weavers understood the threat immediately. There were riots in Lyon, and looms were destroyed. This is where the word Luddite belongs, though the English Luddites of 1811 to 1816 were mostly attacking other machines.

TECHNICAL14.2.5 the engineer’s version#

1. The Jacquard head is a parallel-read, sequential-advance control store. Card width sets the number of parallel control lines.
2. Typical nineteenth-century heads came in sizes such as 400, 600 and 900 hooks. A 900-hook head is a 900-bit-wide control word.
3. Advance is mechanical and unconditional: a four-sided prism rotates one quarter turn per pick, presenting the next card. There is no branch, no conditional and no addressing.
4. In formal terms the Jacquard head is a finite-state sequencer with an external, replaceable program tape, and no conditional transfer. It is not Turing complete and was never meant to be.
5. The honest version: calling a Jacquard card chain “the first program” is a useful teaching device, not a precise claim. It is the first widely used stored, replaceable, machine-readable instruction sequence. It is not a program in the modern sense because it cannot branch or loop.
6. The lineage is nonetheless direct and documented. Babbage planned card input for the Analytical Engine from 1837, Herman Hollerith used cards for the 1890 census, and IBM’s punched card business ran on that idea into the 1970s.

WORDS14.2.6 remember these#

1. Punched card — stiff card with holes meaning yes or no — a parallel-read binary storage medium.
2. Warp — the threads held tight lengthwise on a loom — the addressable lines the control head selects.
3. Jacquard head — the card-reading box on top of a loom, 1804 — a wide parallel control-store sequencer.
4. Program — a list of instructions a machine follows — an encoded sequence of operations held separately from the mechanism.
5. Branch — choosing what to do next — a conditional transfer of control, which the Jacquard head lacks.

14.3 Charles Babbage and Ada Lovelace#

PLAIN14.3.1 in simple words#

1. In the early 1800s, sailors, bankers and astronomers all used printed books of numbers called tables.
2. The tables were worked out by hand by people whose job title was computer. That word meant a person until the 1940s.
3. People make mistakes. Printers make more. A wrong number in a navigation table could put a ship on rocks.
4. Charles Babbage, an English mathematician, wanted to remove the humans from both steps: calculate by machine, and print by machine, so no typesetter could introduce an error.
5. On 14 June 1822 he described his Difference Engine to the Royal Astronomical Society.
6. It could not multiply. It only added. That was enough, because of a mathematical trick explained in the worked example below.
7. The British government funded it from 1823. By 1842 it had spent over 17,000 pounds, a huge sum, and the machine was not finished.
8. Meanwhile Babbage had a much bigger idea. From 1837 he designed the Analytical Engine, a general-purpose machine.
9. It had a mill to do the arithmetic, a store to hold numbers, a card reader for instructions, and a printer for output.
10. Those four parts map exactly onto a modern computer: processor, memory, input, output.
11. Neither machine was completed in his lifetime. Babbage died in 1871.

PLAIN14.3.2 a picture in your head#

1. Think of a very large mechanical cash register, the size of a small room, made of brass and iron, turned by a handle or a steam engine.
2. Inside are columns of numbered wheels. Each column holds one number.
3. Turning the handle makes the columns add themselves together in a fixed pattern, and a new answer appears at the front.
4. Keep turning and a whole table of numbers pours out, and a printing plate is stamped at the same time.
5. That is the Difference Engine: one fixed pattern, endlessly repeated.
6. Now imagine a second machine where a stack of cards tells the columns what to do this time, and where the machine can look at a result and choose a different card next. That is the Analytical Engine.

Where this comparison breaks: a cash register has one accumulator. Babbage’s store was designed for 1,000 numbers of 40 decimal digits each, which is about 16.6 kilobytes. And no cash register can change its own next action based on a result. That ability to decide is what makes the Analytical Engine a computer and the Difference Engine a calculator.

PLAIN14.3.3 a worked example#

1. Here is the trick that lets a machine that can only add produce a table of squares.
2. Write the squares: 1, 4, 9, 16, 25, 36.
3. Subtract each from the next. These are the first differences: 3, 5, 7, 9,
4. Subtract those from each other. The second differences: 2, 2, 2, 2. Constant.
5. So you can run the process backwards using only addition.

6. Each row: add the second difference to the first difference, then add the first difference to the value. Two additions per table entry.
7. Any polynomial of degree n has constant nth differences, so this works for a whole family of useful functions.
8. Logarithms and sines are not polynomials, but they can be approximated by polynomials over short ranges, which is how real tables were produced.
9. Difference Engine No. 1 was designed for 20-digit numbers and sixth-order differences. Difference Engine No. 2, designed between 1847 and 1849, handled 31-digit numbers and seventh-order differences.

PLAIN14.3.4 what is really happening inside#

1. Now Ada Lovelace. She was born in 1815, the daughter of the poet Lord Byron, and was taught mathematics from childhood.
2. She met Babbage in 1833 and followed the Analytical Engine work closely.
3. In 1840 Babbage lectured in Turin. An Italian engineer, Luigi Federico Menabrea, wrote the lecture up in French and published it in 1842.
4. Lovelace translated Menabrea’s article into English over about nine months in 1842 and 1843, and added seven notes of her own, labelled A to G.
5. Her notes are roughly three times longer than the article she translated.
6. Note G contains a step-by-step table showing how the Analytical Engine would compute the Bernoulli numbers, a sequence used in advanced arithmetic.
7. That table has variables, operations, and a loop. It is widely called the first published computer program.
8. Note A contains something arguably more important. Lovelace saw that the numbers in the machine need not stand for quantities at all.
9. She wrote that if the relations of pitched sounds could be expressed as numbers, the engine might compose music. That is the idea of general computing, written down in 1843.
10. She also wrote that the machine has no pretensions to originate anything, and can only do what we know how to order it to perform. That line is still quoted in arguments about artificial intelligence today.

TECHNICAL14.3.5 the engineer’s version#

1. Analytical Engine specification, as designed from 1837: store of 1,000 numbers of 40 decimal digits; decimal, not binary; mill performing the four operations; separate operation cards and variable cards; conditional branching via a mechanism Babbage called the “combinatorial” or barrel control; output to printer, curve plotter and bell.
2. Estimated speed, from Menabrea’s account of Babbage’s claims: an addition in about one second, and the product of two 20-digit numbers in about three minutes.
3. Babbage anticipated microprogramming: the barrel held sequences of micro-operations that a single instruction card triggered. Maurice Wilkes, who invented microprogramming in 1951, noted the resemblance.
4. The scholarly argument about Lovelace, stated fairly, both sides:
   1. Against: Allan Bromley wrote in 1990 that all but one of the programs in her notes had been prepared by Babbage three to seven years earlier. Bruce Collier, also in 1990, judged that she publicized rather than advanced the design. Eugene Eric Kim and Betty Alexandra Toole argued in 1999 that calling her the first programmer is incorrect.
   2. For: Doron Swade, who led the Science Museum’s Babbage work, holds that she alone saw that numbers could represent things other than quantity, which is the conceptual leap to general-purpose computing. Stephen Wolfram, writing in 2016, judged her Bernoulli computation as sophisticated as anything Babbage produced.
5. The defensible statement: Babbage wrote earlier program-like traces that were not published; Lovelace produced the first published, fully worked algorithm intended for a machine, and the clearest statement of what a general-purpose computer would mean. Both facts are true at once.
6. Neither engine was finished, for three reasons: cost, Babbage’s 1833 quarrel with his engineer Joseph Clement over tool ownership, and Babbage’s habit of redesigning before finishing.
7. The vindication came late. The Science Museum in London built Difference Engine No. 2 to the original 1847 to 1849 drawings, completing the calculating section in 1991 for Babbage’s 200th birthday, and the printer in
   2002. It has about 8,000 parts and weighs around 5 tonnes. It works.

WORDS14.3.6 remember these#

1. Difference Engine — Babbage’s table-printing adder, 1822 — a fixed-function polynomial evaluator using finite differences.
2. Analytical Engine — Babbage’s general machine, 1837 — a programmable decimal computer with separate mill, store and card input.
3. Mill — the part that does the sums — the arithmetic and logic unit.
4. Store — the part that remembers numbers — main memory.
5. Method of differences — building a table by adding only — evaluating a degree-n polynomial from constant nth differences.
6. Note G — Lovelace’s Bernoulli number table, 1843 — the first published algorithm written for a machine.

14.4 Hollerith, the 1890 census, and the birth of IBM#

PLAIN14.4.1 in simple words#

1. The United States counts its whole population every ten years. The constitution requires it.
2. The 1880 count took about eight years to process by hand.
3. The country was growing so fast that the 1890 count risked not being finished before the 1900 count began. That is a real crisis, not a joke.
4. Herman Hollerith, a young engineer who had worked on the 1880 census, had an idea borrowed from railway tickets and from Jacquard’s loom.
5. Punch each person’s answers as holes in a card. One card, one person.
6. Then build a machine that reads the holes electrically and counts them.
7. He was granted his key United States patent, number 395,782, on 8 January
8. His machines were used for the 1890 census. The main population count was announced within about six weeks, and the full processing took roughly six years instead of eight, on a much larger and more detailed job.
9. Hollerith founded the Tabulating Machine Company in 1896.
10. In 1911 it merged with three other firms to form the Computing-Tabulating-Recording Company, or CTR.
11. In 1924, under Thomas J. Watson, CTR was renamed International Business Machines. That is IBM.

PLAIN14.4.2 a picture in your head#

1. Imagine a hotel with 200 rooms and a wall of pigeonholes behind reception.
2. Every guest hands in a card with holes punched in it: nationality, age group, room type.
3. The receptionist drops each card into a slot. A lid comes down. Little metal pins press against the card.
4. Where a hole is, a pin dips through and touches mercury in a cup underneath. That completes an electric circuit and a dial clicks up by one.
5. At the same moment, a lid pops open on one pigeonhole, telling the clerk where to file that card.
6. A hundred dials, a hundred questions, one card at a time, all day.

Where this comparison breaks: the real machine was not automatic in the way we would expect. An operator fed every card by hand. The sorting box only opened the correct lid; the human still moved the card. Fully automatic card feeding came later, with Hollerith’s own 1900 machines and their successors.

PLAIN14.4.3 a worked example#

1. Say we want the count of people in a city who are both over 60 and born abroad.
2. Each card has a position for “age band 60 plus” and a position for “born abroad”.
3. The tabulator is wired so a counter advances only when both pins make contact at once.
4. That wiring is a physical AND gate, built from relays, decades before anyone used the word.

Card position 34 punched  -> pin 34 dips -> circuit A closed
Card position 61 punched  -> pin 61 dips -> circuit B closed

Counter 7 advances only if A AND B are closed.

5. Change the question by rewiring the plugboard, not by rebuilding the machine.
6. Speed in practice: a trained operator could process roughly 80 cards a minute on the 1890 equipment, so about 1,000 cards an hour with breaks.
7. Roughly 62.9 million people were counted in the 1890 census. The scale is what made the machinery pay for itself.

PLAIN14.4.4 what is really happening inside#

1. The card is the data. The machine is a fixed reader. The plugboard is the program.
2. This is the same split as the Jacquard loom, but now the cards hold facts rather than instructions, and a separate wiring panel holds the instructions.
3. Hollerith’s business insight mattered as much as his engineering. He rented the machines rather than selling them, and he sold the blank cards.
4. That model, cheap machine, endless consumable, is the razor-and-blades business. It funded IBM for sixty years.
5. IBM standardized an 80-column rectangular-hole card in 1928. The 80-column card is a genuine industry standard, and it is why terminal screens were 80 characters wide for decades, and why many text tools still default to 80.
6. There is a dark chapter that honesty requires naming. IBM’s German subsidiary Dehomag supplied punched card equipment used by the Nazi state in the 1930s and 1940s, including for censuses and record-keeping. Historian Edwin Black documented this in his 2001 book IBM and the Holocaust. The degree of knowledge and control at IBM headquarters is disputed by historians; the supply of equipment is not.

TECHNICAL14.4.5 the engineer’s version#

1. The 1890 card was about 3.25 by 7.375 inches, chosen to match the United States dollar bill of the day so existing cash drawers could store it.
2. The 1890 census card carried up to 288 punch positions in a 12 by 24 grid.
3. Reading was electromechanical: a spring-loaded pin array descends; a pin passing through a hole contacts a mercury cup, closing a circuit that steps a clockwork-style counter dial reading 0 to 9,999.
4. The sorter box had 26 compartments with electrically released lids.
5. The IBM card of 1928 has 80 columns by 12 rows, rectangular holes, and encodes one character per column using Hollerith code: a zone punch in rows 12, 11 or 0 combined with a digit punch in rows 1 to 9.
6. The card measures 7.375 by 3.25 inches by 0.007 inches thick. That thickness is why a box of 2,000 cards is about 14 inches tall.

7. Card equipment was not a dead end. Unit record machines, sorters, collators, reproducers and accounting machines, ran commercial data processing until magnetic tape displaced them through the 1960s.
8. The last widespread public use of punched cards was in voting machines. The 2000 United States presidential election recount in Florida turned on partly detached card chads, which is why the word “chad” briefly became famous.

WORDS14.4.6 remember these#

1. Tabulator — a machine that counts punched cards — an electromechanical counting unit driven by card-hole contacts.
2. Plugboard — a wiring panel that sets the question — a removable control panel defining data paths, the program of a unit record machine.
3. Unit record — one card, one thing — the data model where each card is a complete record.
4. Hollerith code — the hole pattern for a character — zone plus digit punch encoding on an 80-column card.
5. CTR — the 1911 merger — Computing-Tabulating-Recording Company, renamed IBM in 1924.
6. Chad — the bit punched out of a card — the removed rectangle of card stock, famous from the 2000 Florida recount.

14.5 The theory arrives: Hilbert, Gödel, Church and Turing#

PLAIN14.5.1 in simple words#

1. Before anyone built a working computer, mathematicians worked out what a computer could never do. That is a strange and wonderful order of events.
2. Around 1920 the German mathematician David Hilbert proposed a plan for all of mathematics.
3. He wanted a fixed set of starting rules from which every true statement could be proved, with no contradictions, and with a mechanical method for deciding whether any given statement was provable.
4. That last part, the mechanical decision method, had a German name: the Entscheidungsproblem, the decision problem. Hilbert and Wilhelm Ackermann stated it clearly in a 1928 textbook.
5. In 1931 Kurt Gödel, a 25-year-old Austrian, destroyed the first two parts.
6. His incompleteness theorems showed that any consistent set of rules strong enough to describe ordinary arithmetic must contain true statements it cannot prove, and cannot prove its own consistency.
7. The third part, the decision problem, was still open. Two people answered it independently in 1936.
8. Alonzo Church at Princeton invented a symbol-rewriting system called the lambda calculus and used it to prove no such decision method exists.
9. Alan Turing in Cambridge invented an imaginary machine and proved the same thing, in a way that anybody can picture.
10. Turing’s imaginary machine turned out to describe every computer that has ever been built.

PLAIN14.5.2 a picture in your head#

1. Picture an endless paper tape, ruled into squares.
2. Each square is blank or holds one symbol from a small fixed alphabet.
3. A small box sits over one square. The box can read that square, rub it out and write a new symbol, and move one square left or right.
4. The box has a state, which you can think of as a mood. It has a small fixed number of moods, written on a card inside it.
5. The card is a table of rules of the form: if my mood is 3 and I see a 1, then write 0, move left, and change to mood 7.
6. That is the whole machine. Tape, head, moods, rule table.
7. There is nothing else. No screen, no memory chip, no operating system.
8. Turing’s shocking claim is that anything a human being can compute by following a written procedure, this box can compute too.

Where this comparison breaks: the tape is infinite, and no real machine has infinite memory. Real computers are finite-state machines. In practice the tape just needs to be long enough for the problem, so the model is useful anyway. The second break: the model says nothing about speed. A Turing machine takes absurdly many steps for tasks a real processor does in one instruction. It answers “possible or impossible”, never “fast or slow”.

PLAIN14.5.3 a worked example#

1. Here is a complete Turing machine that adds 1 to a binary number.
2. The number is written on the tape, most significant bit on the left. The head starts on the leftmost digit. Blank squares are shown as _.

State  Read  ->  Write  Move  New state
right   0         0      R     right
right   1         1      R     right
right   _         _      L     carry
carry   0         1      L     done
carry   1         0      L     carry
carry   _         1      L     done
done    any      same    -     halt

3. Run it on the tape _1011_, which is 11 in decimal.
4. In state right the head walks over 1, 0, 1, 1 and hits the blank.
5. It steps left into state carry on the last 1. Rule: write 0, move left, still carrying.
6. Next square is 1. Write 0, move left, still carrying.
7. Next square is 0. Write 1, move left, done.
8. The tape now reads _1100_, which is 12. Correct.
9. Six rules and two moods add 1 to any binary number of any length. That is the power of the model in one small picture.

PLAIN14.5.4 what is really happening inside#

1. Turing’s paper is called On Computable Numbers, with an Application to the Entscheidungsproblem. He submitted it on 28 May 1936. It was read to the London Mathematical Society on 12 November 1936 and printed in their Proceedings across 1936 and 1937.
2. Church’s paper An Unsolvable Problem of Elementary Number Theory appeared in the American Journal of Mathematics in April 1936, a few months earlier.
3. Turing added something Church did not have: the universal machine.
4. A universal Turing machine is a Turing machine whose rule table is written to read another machine’s rule table off the tape and imitate it.
5. That is the stored-program computer, described in 1936 as a piece of mathematics, nine years before anybody built one.
6. Now the famous impossibility. Ask: can we write a program that reads any program plus its input, and always answers correctly whether that program will eventually stop or run for ever?
7. Suppose such a checker exists. Build a troublemaker program that runs the checker on itself, and then does the opposite: if the checker says “stops”, the troublemaker loops for ever; if the checker says “loops”, it stops.
8. Now ask the checker about the troublemaker. Either answer makes the checker wrong. So no such checker can exist.
9. That is the halting problem, and it is not a limit of today’s computers. It is a limit of computing itself, for ever.
10. The Church-Turing thesis says that lambda calculus, Turing machines, and every other reasonable definition of “effectively computable” describe exactly the same set of problems.
11. It is a thesis, not a theorem. It cannot be proved, because “what a human could compute with paper and pencil” is not a formal object. Ninety years of failed attempts to beat it are the evidence.

TECHNICAL14.5.5 the engineer’s version#

1. Formally a deterministic single-tape Turing machine is a 7-tuple: states Q, input alphabet, tape alphabet, transition function, start state, accept state, reject state.
2. Turing’s own 1936 proof was about “circle-free” machines and the printing problem, not literally about halting. The name halting problem and the modern formulation are due to Martin Davis, in his 1958 book Computability and Unsolvability.
3. The name Church-Turing thesis was coined by Stephen Kleene in his 1952 book Introduction to Metamathematics.
4. Turing complete means a system can simulate a universal Turing machine, given unbounded storage. Turing equivalent means two systems simulate each other.
5. Turing completeness needs surprisingly little: sequencing, a way to store and read unbounded data, and a conditional branch. Anything with those three is almost always complete.
6. Because the bar is so low, systems become Turing complete by accident. Real documented examples, with attributions:

7. Read that table carefully. The PowerPoint result was presented at SIGBOVIK, a deliberately humorous conference, but the construction is real. The Magic result is a genuine peer-reviewed style construction using legal cards.
8. Practical consequences of undecidability, which working engineers meet:
   1. No compiler can decide in general whether your loop terminates, so termination checkers are conservative and refuse some correct programs.
   2. No antivirus can decide in general whether a program is malicious, which is why detection is heuristic and signature-based.
   3. Rice’s theorem, proved by Henry Gordon Rice in 1951, generalizes this: every non-trivial property of a program’s behaviour is undecidable.
9. The honest version: “undecidable” does not mean “hopeless in practice”. It means no method works for every input. Tools such as static analysers, model checkers and the Coq and Lean proof assistants succeed on huge classes of real programs by giving up on the rest.

WORDS14.5.6 remember these#

1. Entscheidungsproblem — Hilbert’s question of 1928, is there a mechanical test for provability — the decision problem for first-order logic, answered no in 1936.
2. Incompleteness — some true things cannot be proved — Gödel’s 1931 theorems about consistent recursively axiomatized systems containing arithmetic.
3. Turing machine — tape, head, moods, rules — an abstract automaton with unbounded tape and a finite transition function.
4. Universal machine — a machine that imitates any other from its description — the theoretical stored-program computer, Turing 1936.
5. Halting problem — you cannot always tell if a program will stop — the undecidability of the halting set, named by Davis in 1958.
6. Church-Turing thesis — all reasonable definitions of computing agree — an unprovable but unbeaten identification of effective calculability with Turing computability.
7. Turing complete — powerful enough to compute anything computable — able to simulate a universal Turing machine given unbounded memory.

14.6 War machines, 1939 to 1945#

PLAIN14.6.1 in simple words#

1. The Second World War pushed several countries to build calculating machines at a scale nobody would have funded in peacetime.
2. Germany encrypted its radio messages with the Enigma, a typewriter-sized machine with rotating wheels.
3. Press a key and an electrical path runs through the wheels and lights a different letter. The wheels then step, so the next letter is enciphered differently.
4. Polish mathematicians broke into Enigma first. Marian Rejewski worked out the internal wiring mathematically in December 1932.
5. In July 1939, weeks before the invasion, Poland handed everything it knew to British and French intelligence. That gift saved years.
6. Britain’s codebreaking moved to Bletchley Park, a country house north of London, in August 1939.
7. Alan Turing and Gordon Welchman designed the Bombe, a machine that hunted for the day’s Enigma settings.
8. A second, much harder German cipher, called Lorenz by the Germans and Tunny by the British, protected messages between high command and field marshals.
9. To attack Tunny, a Post Office engineer named Tommy Flowers built Colossus, the first large-scale electronic digital machine.
10. At the same time Konrad Zuse in Berlin, John Atanasoff in Iowa and Howard Aiken at Harvard were each building very different machines.

PLAIN14.6.2 a picture in your head#

1. Think of a hotel with a combination lock on the front door, and the combination is changed every midnight.
2. You cannot try all combinations, because there are about 159 million million million of them.
3. But you know one thing: the porter always writes the current date on the pad by the door in the same format.
4. So instead of testing combinations against the lock, you test them against that known scrap of text. Any setting that could not have produced it is thrown away instantly.
5. The Bombe is a machine that throws away wrong settings at high speed. It does not decrypt anything. It hands humans a short list to try by hand.
6. The known scrap of text has a name in this trade: a crib.

Where this comparison breaks: the Bombe did not simply test and compare. It exploited a specific weakness, that Enigma never enciphered a letter as itself, and it used Welchman’s diagonal board to propagate a single contradiction through the whole wiring at once, killing thousands of settings per test rather than one. It was a logic-propagation engine, not a brute-force counter.

PLAIN14.6.3 a worked example#

1. Enigma’s setting count, for the standard army machine of the war years:

2. Multiply those and you get about 1.59 times 10 to the power 20 settings.
3. Written out: roughly 159,000,000,000,000,000,000.
4. Testing one per second would take longer than the age of the universe.
5. Now use the crib. Suppose the intercepted text contains WETTERBERICHT, German for weather report, and you can guess where.
6. Enigma never maps a letter to itself. Slide the crib along the ciphertext. Any position where a letter lines up with itself is impossible, and is dropped without any machine at all.
7. The Bombe then took the surviving positions, wired up the logical chain of letter-to-letter implications, and spun until the electrical current found no contradiction. It stopped, and a human checked the stop.
8. A single Bombe run took around 20 minutes. By May 1945 Britain had about 155 three-rotor Bombes plus later four-rotor machines, running day and night.

PLAIN14.6.4 what is really happening inside#

1. The first British Bombe, code-named Victory, was installed at Bletchley Park on 18 March 1940.
2. The improved version with Gordon Welchman’s diagonal board, called Agnus and then Agnes, ran from 8 August 1940. That addition made the whole approach practical.
3. Bombes were built by the British Tabulating Machine Company at Letchworth, with Harold Keen as chief engineer. About 200 were built in Britain.
4. Be clear on this: the Bombe was electromechanical, not electronic, and it was not a computer. It ran one fixed search. It could not be programmed.
5. Colossus is a different animal. The Lorenz cipher used twelve wheels and was attacked statistically, not by simulation.
6. Bill Tutte worked out the complete internal structure of the Lorenz machine in 1941 and 1942 without ever seeing one, from intercepted traffic alone. It is one of the great feats of the war.
7. Max Newman designed a statistical attack. The first machine to run it, in 1943, was electromechanical and nicknamed Heath Robinson. It was too slow and its paper tapes tore.
8. Tommy Flowers, of the Post Office Research Station at Dollis Hill, argued that valves, what Americans call vacuum tubes, could do the job if you never switched them off. His superiors doubted it. He built it anyway, partly with his own money.
9. Colossus Mark 1, with 1,600 valves, was working at Bletchley Park in January 1944 and attacked its first real message on 5 February 1944.
10. Colossus Mark 2, with 2,400 valves, started work at 8 in the morning on 1 June 1944, five days before the Normandy landings. Ten Colossi were running by the end of the war.
11. Colossus was programmable only by switches and plugs, and it had no stored program and no general arithmetic. It was a special-purpose electronic machine. That is still a first.
12. Most Colossi were broken up on Churchill’s orders. The work stayed secret until the mid-1970s, which is why the machine is missing from older history books. A working rebuild led by Tony Sale was completed in 2007.

TECHNICAL14.6.5 the engineer’s version#

1. Konrad Zuse completed the Z3 in Berlin and presented it on 12 May 1941.

2. The Z3 used binary floating point with a 14-bit mantissa, 7-bit exponent and sign bit, decades ahead of American practice. Program came from punched 35 mm film. It had no conditional branch.
3. Raúl Rojas showed in 1998 that the Z3 is Turing complete in principle, but only by an artificial construction that computes all branches. That is a theoretical result, not a description of how it was used.
4. The original Z3 was destroyed on 21 December 1943 in an Allied air raid on Berlin. A working replica was built in 1961.
5. The Atanasoff-Berry Computer was built at Iowa State College by John Vincent Atanasoff and Clifford Berry between 1939 and 1942.
6. It used about 280 vacuum tubes for arithmetic, was binary, held 60 numbers of 50 bits in rotating capacitor drums, and did about 30 additions per second.
7. Its regenerative capacitor memory, refreshed each rotation, is the direct ancestor of the refresh cycle in modern DRAM.
8. It solved systems of up to 29 simultaneous linear equations. It was not programmable and it never worked fully automatically; the card output unit was unreliable.
9. On 19 October 1973, in Honeywell v. Sperry Rand, United States District Judge Earl R. Larson invalidated the ENIAC patent, ruling that Eckert and Mauchly “derived that subject matter from one Dr. John Vincent Atanasoff”.
10. The honest version: that ruling is a legal finding about patent validity, not a historical consensus about who invented the computer. Historians still argue about how much Mauchly took from his 1941 visit to Atanasoff.
11. The Harvard Mark I, built by IBM as the Automatic Sequence Controlled Calculator, was presented to Harvard on 7 August 1944.
12. It was 51 feet long, 8 feet high, weighed about 4.7 short tons, and did 3 additions per second, a multiplication in 6 seconds and a division in 15.3 seconds. It was decimal, driven by a 50-foot shaft, and read its program from punched paper tape.
13. Howard Aiken led it. Grace Hopper, Richard Bloch and Robert Campbell were among its first programmers. John von Neumann ran implosion calculations on it for the Manhattan Project from 29 March 1944.

WORDS14.6.6 remember these#

1. Enigma — the German rotor cipher machine — a stepping-rotor polyalphabetic substitution device with a plugboard.
2. Crib — a guessed piece of the plain message — known plaintext used to constrain a key search.
3. Bombe — the Bletchley setting-hunting machine, 1940 — an electromechanical parallel logic engine testing rotor hypotheses for contradiction.
4. Lorenz or Tunny — the high command teleprinter cipher — a 12-wheel additive stream cipher on 5-bit Baudot code.
5. Colossus — the 1943 to 1944 valve machine — the first large-scale electronic digital special-purpose computer, 1,600 then 2,400 valves.
6. Z3 — Zuse’s 1941 relay machine — the first working programmable automatic binary floating-point computer.
7. ABC — the Atanasoff-Berry Computer of 1942 — the first electronic binary digital calculating device, non-programmable, with regenerative memory.

14.7 ENIAC and EDVAC#

PLAIN14.7.1 in simple words#

1. Artillery needs firing tables: for this gun, this shell, this angle, this wind, where does it land.
2. In 1943 the United States Army was drowning in the arithmetic. Each table needed thousands of trajectory calculations.
3. The Army funded John Mauchly and J. Presper Eckert at the Moore School of Electrical Engineering, University of Pennsylvania, to build a fully electronic calculator.
4. The result was ENIAC, the Electronic Numerical Integrator and Computer.
5. It filled a room, used nearly 18,000 vacuum tubes, weighed 30 tons and drew 150 kilowatts.
6. It did 5,000 additions a second, roughly a thousand times faster than the relay machines of the day.
7. It was finished in 1945 and shown to the press on 14 February 1946.
8. It had one enormous flaw: to change the problem, you rewired the machine.
9. That work took days or weeks of moving cables and setting thousands of switches by hand.
10. The people who did that work were six women, and for forty years almost nobody knew their names.

PLAIN14.7.2 a picture in your head#

1. Think of an old telephone exchange, the kind with a wall of sockets and a human operator pushing plugs into them.
2. Now imagine that instead of connecting callers, each plug decides which number flows into which adding unit.
3. To ask the machine a different question you do not type anything. You unplug forty cables and plug them somewhere else, and turn 3,000 switches.
4. Then you spend two days finding the one cable you put in the wrong hole.
5. That is ENIAC. The program was the physical shape of the wiring.

Where this comparison breaks: a telephone exchange connects two points at a time and then releases them. ENIAC’s plugboard connections were the permanent data paths for the whole run, and the timing of pulses down those cables was the control flow. It is closer to laying out a factory production line than to placing a call.

PLAIN14.7.3 a worked example#

1. Here are ENIAC’s real numbers, from the Moore School records.

2. You will often see 18,000 tubes quoted. The precise figure is 17,468. Both appear in reputable sources; the rounder number came from press releases.
3. Arithmetic detail that surprises people: ENIAC was decimal, not binary. It had 20 accumulators, each holding 10 decimal digits and a sign.
4. A digit was stored as a ring of ten flip-flops, one of which was on. That is ten tubes to hold what one bit-pair would hold in binary.
5. Reliability was the real problem. With 17,468 tubes, a failure every day or two was expected. Eckert ran the tubes far below rated voltage and never switched the machine off, since switch-on is when tubes die.
6. ENIAC’s first serious job, in December 1945, was a feasibility calculation for the hydrogen bomb, run for six weeks. Its very first task was not ballistics at all.
7. It was switched off for the last time at 11:45 in the evening on 2 October

PLAIN14.7.4 what is really happening inside#

1. The Army had already hired about eighty women as human computers, working out trajectories on desk calculators.
2. Six of them were selected in 1945 to program ENIAC: Kathleen McNulty, Frances Bilas, Betty Jean Jennings, Ruth Lichterman, Elizabeth Snyder and Marlyn Wescoff.
3. They are also known by their later married names: McNulty Mauchly Antonelli, Bilas Spence, Jennings Bartik, Lichterman Teitelbaum, Snyder Holberton, and Wescoff Meltzer.
4. They were given wiring diagrams and told to work it out. There was no manual, no programming language and no course, because none existed.
5. They learned the machine from its blueprints, invented debugging techniques, and worked out how to break a calculation into parallel units.
6. Betty Snyder Holberton went on to write the first sort-merge generator and worked on the standards for COBOL and Fortran. Kathleen McNulty helped devise subroutine techniques on later machines.
7. On the day ENIAC was shown to the press in February 1946, the men who built the hardware were named. The women were not invited to the celebratory dinner.
8. For decades their photographs were captioned as models posing with the equipment. The phrase used later to describe this is “refrigerator ladies”, because women in technical photographs were assumed to be advertising props.
9. Kathryn Kleiman began tracking them down from 1986. Their story reached a wide audience with the 1997 Women in Technology International Hall of Fame inductions and the 2014 documentary The Computers.
10. This is not a footnote about fairness. Programming as a distinct discipline, separate from building hardware, starts here, and these six people invented much of its practice.

TECHNICAL14.7.5 the engineer’s version#

1. While ENIAC was still being wired, its team already knew rewiring was unacceptable. The successor project was called EDVAC.
2. John von Neumann joined the group as a consultant in 1944. On 30 June 1945 Herman Goldstine distributed a 101-page document titled First Draft of a Report on the EDVAC.
3. It described a machine with five parts: a central arithmetic unit, a central control unit, memory, input and output.
4. Its crucial claim was that instructions and data live in the same read-write memory, so a program is just numbers, and a program can therefore be loaded, changed and even modified by another program.

  +---------+        +-----------------+        +---------+
  |  Input  | -----> |     Memory      | -----> | Output  |
  +---------+        |  data + program |        +---------+
                     +--------+--------+
                              |
                       +------+------+
                       |   Control   |
                       +------+------+
                              |
                       +------+------+
                       |  Arithmetic |
                       +-------------+

5. The naming controversy is real and worth stating precisely.
   1. The document carried only von Neumann’s name. It was a draft circulated for comment, not a finished credited paper.
   2. Eckert and Mauchly said the stored-program concept came out of Moore School group discussions in 1944, before von Neumann arrived, and that he wrote up shared ideas in logical notation.
   3. Because the draft was circulated publicly more than a year before the patent filing, it counted as prior public disclosure. That helped invalidate the ENIAC patent claims later.
   4. Turing’s 1936 universal machine had already stated the idea mathematically, and von Neumann had read Turing’s paper.
6. Many historians therefore prefer the neutral terms “stored-program computer” or “Princeton architecture”. The label “von Neumann architecture” is a convention, and a slightly unfair one.
7. What the name refers to today is a specific structural property: a single address space and a single bus for both instructions and data. The alternative, separate instruction and data memories, is the Harvard architecture, named after the Harvard Mark I.
8. The consequence engineers still fight is the von Neumann bottleneck, a phrase John Backus coined in his 1977 Turing Award lecture: the processor can only go as fast as the single path to memory allows.
9. Nearly every modern chip is a hybrid. Level 1 caches are usually split into separate instruction and data caches, which is Harvard-like, over a unified main memory, which is von Neumann. You can see this on Linux with lscpu, which lists L1d and L1i separately.
10. EDVAC itself was delivered in 1949 and became properly operational in 1951, by which time British machines had beaten it into service.

WORDS14.7.6 remember these#

1. ENIAC — the 1945 American electronic calculator — a decimal, plugboard programmed, 17,468-tube general-purpose electronic computer.
2. Accumulator — a box that holds a running total — a register that stores an arithmetic result, ten decimal digits on ENIAC.
3. Stored program — instructions kept in memory like data — a single writable address space holding both code and operands.
4. von Neumann architecture — the 1945 five-part design — a stored-program machine with unified instruction and data memory.
5. Harvard architecture — separate stores for code and data — physically distinct instruction and data memories and buses.
6. von Neumann bottleneck — the narrow pipe to memory — the throughput limit imposed by a single shared processor-memory path, named by Backus in 1977.

14.8 The first stored-program computers#

PLAIN14.8.1 in simple words#

1. After 1945 the race was to build a machine that kept its program in memory.
2. The hard part was not logic. It was memory. Nobody had a fast, cheap, rewritable store.
3. Two solutions appeared. One used sound. One used a television tube.
4. The sound one is the mercury delay line. Send a pulse of sound down a tube of mercury, catch it at the far end, and send it round again. The data is literally in flight.
5. The television one is the Williams tube. Draw dots on the inside of a cathode ray tube. The charge left behind is the bit. Read it with a metal plate on the glass.
6. At Manchester, Freddie Williams and Tom Kilburn made the tube work, and then built the smallest possible computer to prove it.
7. That machine, the Small-Scale Experimental Machine, nicknamed the Baby, ran the world’s first stored program on 21 June 1948.
8. At Cambridge, Maurice Wilkes built EDSAC, which first ran on 6 May 1949 and was the first machine to give a regular computing service to other people.
9. Manchester’s full-size machine became the Manchester Mark 1, and the company Ferranti turned it into a product.
10. The Ferranti Mark 1, delivered in February 1951, was the first commercially available general-purpose electronic computer in the world.
11. In America, UNIVAC I was accepted by the Census Bureau on 31 March 1951 and became famous on election night in 1952.

PLAIN14.8.2 a picture in your head#

1. Picture a long queue of people passing a whispered message down the line and back to the front, over and over, so it is never forgotten.
2. That is a mercury delay line. The message exists only as sound travelling through liquid metal.
3. To read a particular bit, you wait for it to come round. If you just missed it, you wait a full lap.
4. Now picture instead a chalk mark on a blackboard that fades in a fraction of a second, so a helper walks along constantly redrawing every mark.
5. That is a Williams tube. Any dot you can see, you can read at once, in any order.

Where this comparison breaks: the queue is not just slow, it is strictly sequential, which changed how programs were written. On EDSAC, and even more on drum-memory machines, programmers placed instructions at positions on the loop so the next one arrived just as it was needed. That practice is called optimum coding, and it is the ancestor of today’s cache-aware programming.

PLAIN14.8.3 a worked example#

1. The Baby’s first program, written by Tom Kilburn, found the highest proper factor of 2 to the power 18, which is 262,144.
2. The method was the crudest possible: try every candidate downwards from 262,143 and test by repeated subtraction.
3. It was 17 instructions long and it lived in a Williams tube holding 32 words of 32 bits, which is 1,024 bits in total.
4. It ran for about 52 minutes and performed roughly 3.5 million operations, which is around 1,100 operations per second.
5. The answer, of course, is 131,072. The point was never the answer. The point was that the program was in memory.
6. Compare that with EDSAC’s first run on 6 May 1949, which printed a table of the squares of 0 to 99, and then a list of prime numbers.

PLAIN14.8.4 what is really happening inside#

1. EDSAC’s real importance is not the hardware. It is that Cambridge treated computing as a service and a discipline.
2. David Wheeler invented the subroutine there: a block of instructions you can call from anywhere and return from. The return mechanism is still called the Wheeler jump.
3. Wilkes, Wheeler and Gill published The Preparation of Programs for an Electronic Digital Computer in 1951. It is the first programming textbook, and it introduced the idea of a library of reusable routines.
4. Wheeler also received what is generally described as the world’s first doctorate in computer science, in 1951.
5. EDSAC also produced the first business computing. The catering company J. Lyons and Company built a copy called LEO, the Lyons Electronic Office, which ran its first routine business job on 17 November 1951, valuing a bakery’s output.
6. A tea shop company thus ran payroll on a computer before most governments did. That is a genuine and under-told fact.
7. The Manchester Mark 1 contributed the index register, which Manchester called a B-line: a register whose value is added to an address before use.
8. That one idea makes loops over arrays possible without a program rewriting its own instructions, and every processor since has it.

TECHNICAL14.8.5 the engineer’s version#

1. Williams-Kilburn tube: a standard CRT storing charge on the phosphor. Typical capacity 1,024 to 2,048 bits per tube. Access time about 10 to 30 microseconds and, crucially, random access. It needed constant refresh and was sensitive to nearby electrical noise, even to someone opening a door.
2. Mercury delay line: a tube roughly 1.5 metres long. Sound travels in mercury at about 1,450 metres per second, giving a loop time near 1 millisecond and holding several hundred bits. Average latency is half a loop.
3. EDSAC: 3,000 valves, about 11 kW, 512 words of 17 bits initially, cycle time 1.5 ms for most instructions and 6 ms for multiplication. It ran until 11 July 1958.
4. Ferranti Mark 1: delivered to the University of Manchester in February 1951, publicly demonstrated in July 1951. Around nine machines including the improved Mark 1 star were built between 1951 and 1957.
5. Priority claims, stated carefully, because people argue about them:
   1. First stored program executed: Manchester Baby, 21 June 1948.
   2. First regular computing service: EDSAC, from 1949.
   3. First commercially available general-purpose computer delivered: Ferranti Mark 1, February 1951.
   4. First American commercial delivery: UNIVAC I, accepted by the Census Bureau on 31 March 1951, physically installed later that year.
   5. Zuse’s Z4 was delivered to ETH Zurich in July 1950 and was electromechanical, so it is usually excluded from electronic firsts.
6. UNIVAC I specifications: 5,200 vacuum tubes, about 125 kW, 1,000 words of 12 characters in mercury memory, about 1,905 operations per second, 46 machines built, and a price that rose from a quoted $159,000 to well over a million dollars.
7. The 1952 election. On CBS, on 4 November 1952, UNIVAC was given 5.5 per cent of the returns and predicted a landslide for Dwight Eisenhower over Adlai Stevenson, at 438 electoral votes to 93.
8. Nobody believed it, because pollsters expected a close race. CBS staff adjusted a trend factor from 40 per cent to 4 per cent to produce a respectable-looking 268 to 263, and broadcast that instead.
9. The actual result was 442 to 89. The original prediction was within four electoral votes. Late that night Remington Rand’s Arthur Draper admitted on air that the machine had been right and had been overruled.
10. That broadcast, more than any technical paper, is what put the word “computer” into ordinary speech.

WORDS14.8.6 remember these#

1. Mercury delay line — memory made of sound in a tube — a sequential-access acoustic recirculating store, about 1 ms loop time.
2. Williams tube — memory made of dots on a screen — a CRT charge-storage random-access memory, about 1 to 2 kilobits per tube.
3. Baby — the Manchester test machine of 1948 — the Small-Scale Experimental Machine, first to execute a program held in memory, 21 June 1948.
4. Subroutine — a reusable block of instructions — a callable routine with a return mechanism, invented by David Wheeler on EDSAC.
5. Index register — a register that shifts an address — a B-line, added to an operand address before memory access, from the Manchester Mark 1.
6. LEO — the tea company’s computer, 1951 — Lyons Electronic Office, the first computer used for routine commercial work.

14.9 Transistors, mainframes and minicomputers#

PLAIN14.9.1 in simple words#

1. A vacuum tube is a small glass bulb that controls electricity. It works, but it is hot, fragile, power-hungry and it burns out.
2. On 23 December 1947, John Bardeen and Walter Brattain at Bell Labs demonstrated the transistor: a lump of germanium doing the same job with no glass, no heater and almost no power.
3. William Shockley followed with the more manufacturable junction transistor, published in 1951. The three shared the Nobel Prize in Physics in 1956.
4. Machines built from transistors were smaller, cooler and ran for months instead of hours between failures.
5. Through the 1950s IBM sold its 700 series, built from tubes, and then the 7000 series, built from transistors.
6. Every one of those machines had its own instruction set, so software written for one would not run on the next.
7. On 7 April 1964 IBM announced the System/360: one family of machines, from small to huge, all running the same programs.
8. That single decision made software an asset you keep instead of a cost you repeat, and it made IBM dominant for twenty years.
9. Meanwhile a small company called Digital Equipment Corporation went the other way, building machines cheap enough for one laboratory to own.
10. Those were minicomputers, and they took computing out of the glass room and put it next to the people doing the work.

PLAIN14.9.2 a picture in your head#

1. A mainframe is a city power station. It is enormous, it is behind a fence, and you apply for a connection.
2. Programs arrive as decks of cards, are run overnight in a queue, and results come back as printout the next morning. That is batch processing.
3. A minicomputer is a generator in your own workshop. Smaller, noisier, less efficient, but yours, and switched on when you want it.
4. Nobody schedules you. You sit at the machine, you type, it answers.
5. System/360 is a third idea: not one power station but a whole product range, from a small generator to a station, all taking the same plug.

Where this comparison breaks: minicomputers were not personal. A PDP-11 cost tens of thousands of dollars and served a whole department, usually through several terminals at once. The truly personal machine is still a decade away in this story.

PLAIN14.9.3 a worked example#

1. Here is why compatibility mattered, in money.
2. Say a bank has 200 programs, each taking 6 months of one programmer’s time to write. That is 100 programmer-years of investment.
3. Before 1964, buying a bigger machine meant rewriting most of that. The software bill could exceed the hardware bill.
4. With System/360, the same program ran on a Model 30 and on a Model 75, which was roughly fifty times faster.
5. You bought a bigger box and your software simply ran faster. Nothing else changed.

PLAIN14.9.4 what is really happening inside#

1. The trick that made one family possible is microprogramming, invented by Maurice Wilkes at Cambridge in 1951.
2. Instead of building each instruction directly out of logic gates, you build a small, fast inner machine, and store a tiny program for each instruction in read-only memory.
3. A cheap model runs the same instruction set with slow, simple hardware and more microcode steps. An expensive model uses fast hardware and fewer steps.
4. Same behaviour, wildly different cost. That is how one architecture spanned a fifty-to-one performance range.
5. Digital Equipment Corporation, founded in 1957 by Ken Olsen and Harlan Anderson, attacked from below.
6. The PDP-1 of 1959 cost $120,000 when mainframes cost millions. Fifty-three were built. On one of them, in 1962, Steve Russell and friends at MIT wrote the game Spacewar, one of the first video games.
7. The PDP-8, launched on 22 March 1965 at $18,500, was the first computer sold for under $20,000. Over 50,000 were sold, and more than 300,000 across the whole family.
8. The PDP-11, announced in January 1970, sold around 600,000 units over its life. Unix was first written for it, and the C language grew up on it.

TECHNICAL14.9.5 the engineer’s version#

1. System/360 announcement, 7 April 1964: six models initially, numbered 30, 40, 50, 60, 62 and 70. The last three were replaced before shipping by the 65 and 75.
2. Development cost is usually quoted as about $5 billion, from a September 1966 Fortune article calling it IBM’s five-billion-dollar gamble. Adjusted for inflation that is roughly $50 billion today.
3. Architectural decisions from System/360 that are still with us:
   1. The 8-bit byte as the standard addressable unit. Before 360, machines used 6-bit characters and word sizes of 36, 48 or 60 bits.
   2. 32-bit words, 24-bit addressing, and general-purpose registers.
   3. A separate channel subsystem for input and output, which is the ancestor of DMA controllers.
4. Gene Amdahl was chief architect. Fred Brooks managed the OS/360 software project, and later wrote The Mythical Man-Month in 1975 about how badly it went. Its most quoted line: adding people to a late software project makes it later.
5. IBM’s dominance drew the United States Department of Justice, which filed an antitrust suit on 17 January 1969. It ran for thirteen years and was dropped on 8 January 1982 as being without merit.
6. The rest of the industry was known as the BUNCH: Burroughs, UNIVAC, NCR, Control Data and Honeywell. The older joke was “IBM and the seven dwarfs”.
7. Mainframes did not die. IBM z16 systems, announced in April 2022, still run code compiled for System/360 in 1964. That is 60 years of binary compatibility, the longest unbroken run in the industry.

WORDS14.9.6 remember these#

1. Transistor — a solid switch with no glass bulb — a semiconductor device controlling current, demonstrated at Bell Labs in December 1947.
2. Mainframe — a big shared central computer — a high-throughput, high-reliability system optimized for input and output volume.
3. Minicomputer — a machine a department can afford — a mid-range system, typically 12 to 16 bits, priced from about $18,500 in 1965.
4. Batch processing — hand in your job, come back tomorrow — non-interactive scheduled execution of queued jobs.
5. Microprogram — a tiny program inside the processor — microcode in control store implementing each machine instruction, invented by Wilkes in 1951.
6. Binary compatibility — old programs still run — the ability to execute unmodified object code on a newer implementation.

14.10 The integrated circuit and Silicon Valley#

PLAIN14.10.1 in simple words#

1. By 1958 a computer was thousands of separate parts joined by hand-soldered wires. Every joint was a chance to fail.
2. Engineers called this the tyranny of numbers: to make a machine more capable you needed more parts, and more parts meant it broke sooner.
3. On 12 September 1958, Jack Kilby at Texas Instruments demonstrated a different idea: build all the parts out of one piece of semiconductor, so there are no joints at all.
4. His device worked but used tiny gold wires bonded by hand across the surface. It was a proof, not a product.
5. In 1959 Robert Noyce at Fairchild Semiconductor solved the manufacturing problem, using a silicon wafer and printing the connections on as a flat layer of aluminium.
6. That is the integrated circuit, or chip.
7. Both men are credited. Kilby received the Nobel Prize in Physics in 2000; Noyce had died in 1990 and Nobel prizes are not awarded posthumously.
8. Noyce’s version won commercially because it could be made in quantity.
9. The reason it could be made in quantity was a process invented by his colleague Jean Hoerni, also in 1959, called the planar process.
10. All of this happened in a strip of orchards south of San Francisco, which a journalist named Silicon Valley in 1971.

PLAIN14.10.2 a picture in your head#

1. Think of printing a newspaper instead of writing each copy by hand.
2. Old way: draw every letter individually, then glue the pages together.
3. New way: make one master plate, then print thousands of identical copies at almost no extra cost per copy.
4. A chip is printed. Light is shone through a mask onto a light-sensitive coating on the silicon, and the pattern is etched in.
5. Printing 10 components costs the same as printing 10 million. That is the entire economics of the computer industry in one sentence.

Where this comparison breaks: newspapers do not care about a single dust speck. A chip does. One particle in the wrong place kills the die, so yield, the fraction of good chips per wafer, dominates cost. That is why chips are made in cleanrooms with fewer particles per cubic metre than an operating theatre.

PLAIN14.10.3 a worked example#

1. The planar process in five steps, which is still roughly what happens today.

1. Oxidize   : grow a glass layer of SiO2 on the silicon
2. Mask      : coat with resist, shine light through a mask
3. Etch      : dissolve the exposed oxide, opening windows
4. Diffuse   : drive dopant atoms in through the windows
5. Metallize : evaporate aluminium on top, then pattern it

2. The key point is step 1. The oxide stays on as a permanent protective and insulating layer.
3. Because the surface is flat and insulated, metal wiring can be laid across the top without shorting to anything underneath.
4. Kilby’s germanium device had no such layer, so his connections were flying wires. Hoerni’s flat oxide is what made mass production possible.
5. Kurt Lehovec at Sprague Electric added the other missing piece in 1959: isolating neighbouring components with reverse-biased junctions.

PLAIN14.10.4 what is really happening inside#

1. William Shockley left Bell Labs and set up Shockley Semiconductor Laboratory in Mountain View, California, in 1956, the year he won the Nobel Prize.
2. He hired brilliant people and managed them badly. He posted salaries publicly and required staff to take lie detector tests.
3. On 18 September 1957 eight of them resigned together. Shockley called them the traitorous eight.
4. They were Julius Blank, Victor Grinich, Jean Hoerni, Eugene Kleiner, Jay Last, Gordon Moore, Robert Noyce and Sheldon Roberts.
5. Sherman Fairchild financed them with a loan of $1.38 million, and Fairchild Semiconductor was born.
6. Fairchild then leaked people the way Shockley had. Companies founded by ex-Fairchild staff are called Fairchildren, and there are dozens.
7. On 18 July 1968 Noyce and Moore left to found Intel, backed by the venture capitalist Arthur Rock.
8. On 1 May 1969 Jerry Sanders, Fairchild’s head of marketing, founded Advanced Micro Devices with seven colleagues.
9. Eugene Kleiner went on to co-found the venture firm Kleiner Perkins in 1972, which is how the engineering diaspora became a financial one too.
10. The name arrived on 11 January 1971, when the journalist Don Hoefler ran a series titled “Silicon Valley U.S.A.” in the trade paper Electronic News.
11. Silicon is the element the chips are made from; the valley is the Santa Clara Valley. The name did not become common speech until the early 1980s.

TECHNICAL14.10.5 the engineer’s version#

1. Kilby’s 12 September 1958 device: a phase-shift oscillator on a germanium bar about 11 by 1.6 millimetres, with gold wire interconnect. Texas Instruments filed the patent in February 1959.
2. Noyce’s Fairchild device: silicon, oxide-passivated by Hoerni’s planar process, with evaporated aluminium interconnect. Patent filed 30 July 1959.
3. The two companies litigated for a decade and settled with cross-licensing in
   1966. Courts eventually favoured Noyce on the interconnect claims and Kilby on the integration concept.
4. Gordon Moore’s observation was published in Electronics magazine on 19 April
   1965. He noted the number of components per chip had roughly doubled each year, and predicted it would continue for a decade.
5. In 1975 Moore revised the doubling period to about two years. The often-quoted eighteen months is a later paraphrase attributed to Intel colleague David House, combining transistor count and speed.
6. Standard, convention or observation? Moore’s law is none of a standard, a law of physics or a theorem. It is an economic observation that became a planning target the industry then organized itself to meet.

WORDS14.10.6 remember these#

1. Integrated circuit — many parts on one piece of silicon — a monolithic circuit fabricated on a single semiconductor die.
2. Planar process — the flat oxide-layer way of making chips — Hoerni’s 1959 oxide passivation and diffusion method enabling deposited interconnect.
3. Wafer — the silicon disc chips are printed on — a monocrystalline substrate, today 300 mm in diameter.
4. Yield — how many chips per wafer actually work — the ratio of functional die to total die, the dominant cost factor.
5. Moore’s law — chips double in capacity every couple of years — the 1965 observation on component count per die, revised to two years in 1975.
6. Traitorous eight — the group who quit Shockley in 1957 — the founders of Fairchild Semiconductor and, indirectly, of Silicon Valley.

14.11 The microprocessor#

PLAIN14.11.1 in simple words#

1. By 1969 you could put thousands of transistors on one chip, but a computer’s processor was still spread over many chips and several boards.
2. In April 1969 a Japanese calculator company called Busicom asked Intel, then a young memory company, for a set of chips for a printing calculator.
3. Busicom’s own design used twelve custom chips, each hardwired for one part of the job.
4. An Intel engineer, Ted Hoff, said that was backwards. Build one small general-purpose processor, and put the calculator’s behaviour in a program in read-only memory.
5. Stanley Mazor helped work out the instruction set. Busicom’s engineer Masatoshi Shima wrote much of the calculator software and checked the design.
6. Nobody at Intel could actually build it until Federico Faggin joined in April 1970 and invented the circuit techniques needed to fit it on one die.
7. The result was the Intel 4004, announced on 15 November 1971.
8. It had 2,300 transistors, a 4-bit word, ran at up to 740 kilohertz, and cost $60.
9. That single chip had roughly the computing power of ENIAC, which had filled a room in 1946.
10. Intel then bought back the exclusive rights from Busicom, which was in financial trouble, by refunding $60,000 of development cost. That is how a calculator part became a product anyone could buy.

PLAIN14.11.2 a picture in your head#

1. Think of a kitchen that makes one dish. Every worktop, jig and mould is shaped for that dish alone. Change the menu and you rebuild the kitchen.
2. Now think of a kitchen with one general chef, a knife, a pan and a recipe card.
3. The second kitchen is slower at the one dish, but it can cook anything, and changing the menu costs the price of a card.
4. The 4004 is the general chef. The read-only memory chip is the recipe card.
5. Because one design serves a thousand products, you can afford to make ten million of them, and the price collapses.

Where this comparison breaks: general-purpose chips did not simply beat custom ones. Custom silicon is still faster and cheaper per unit for very high volume fixed jobs, which is why your phone contains dozens of fixed-function blocks for video, radio and camera work alongside its general processors.

PLAIN14.11.3 a worked example#

1. Compare the 4004 with ENIAC, which is a fair fight on arithmetic rate.

2. Twenty-five years took a room-sized machine to a fingernail-sized chip, at about one eight-thousandth of the price.
3. The 4004 was made on a 10 micrometre pMOS silicon-gate process. Today’s leading process is around 3 nanometres, which is roughly 3,000 times finer in each direction.

PLAIN14.11.4 what is really happening inside#

1. The 4004 was a starting gun, not a finish line. The useful chips came next.
2. The Intel 8008 arrived in April 1972. It was 8-bit and came from a separate contract for the Datapoint 2200 terminal, whose maker rejected it.
3. The Intel 8080 arrived in April 1974, with about 4,500 transistors and a 2 megahertz clock. It was fast enough to be the heart of a real computer, and it powered the Altair 8800.
4. Motorola answered with the 6800 in 1974. It needed only a single 5-volt supply, which made board design far simpler.
5. Federico Faggin left Intel at the end of 1974 and founded Zilog. With Masatoshi Shima he designed the Z80, released in July 1976.
6. The Z80 ran all 8080 software, added instructions and registers, needed one power supply, and refreshed dynamic memory itself. It became the most widely used 8-bit processor of the era.
7. The turning point for ordinary people was price. In August 1974 the Motorola 6800 and Intel 8080 were introduced at around $360 each.
8. Chuck Peddle had worked on the 6800 at Motorola. In August 1974 he and seven colleagues left for MOS Technology.
9. In September 1975, at the Wescon show in San Francisco, they sold the MOS 6502 for $25, out of jars on a table, because nobody believed the price otherwise.
10. Twenty-five dollars instead of three hundred and sixty is what made a home computer possible. The 6502 went into the Apple I and II, the Commodore PET and 64, the BBC Micro, the Atari 2600 and the Nintendo Entertainment System.

TECHNICAL14.11.5 the engineer’s version#

1. Faggin’s enabling technique was silicon gate MOS, which he had developed at Fairchild in 1968. Using polysilicon rather than aluminium for the gate makes it self-aligning, which shrinks parasitic capacitance and allows buried contacts. Without it the 4004 would not have fitted on one die.
2. The 4004 shipped as part of the MCS-4 chip set: 4001 ROM, 4002 RAM, 4003 shift register and 4004 CPU. On its own the processor is not a computer.
3. It used a 4-bit data bus, a 12-bit address space of 4 kilobytes of program memory, 46 instructions, and a 16-level address stack in hardware.

4. The 6502 was cheap for two engineering reasons, not just aggressive pricing. First, MOS Technology fixed mask defects rather than discarding the wafer, raising yield dramatically. Second, the design is minimal: three 8-bit registers, a hardware stack fixed at page one, and no microcode.
5. Its zero-page addressing mode treats the first 256 bytes of memory as quasi-registers, which is how a chip with so few registers stayed fast.
6. Legal note: MOS Technology first shipped the 6501, which was pin-compatible with the Motorola 6800. Motorola sued in November 1975. MOS withdrew the 6501 and settled in 1976, paying $200,000. The 6502 was not pin-compatible and survived.

WORDS14.11.6 remember these#

1. Microprocessor — a whole processor on one chip — a monolithic CPU, first commercially released as the Intel 4004 in November 1971.
2. Silicon gate — a way of building transistors with polysilicon gates — self-aligned MOS process reducing gate capacitance, Faggin 1968.
3. Instruction set — the list of things a chip understands — the architectural contract between hardware and compiled code.
4. Read-only memory — the chip holding the fixed program — ROM, non-volatile storage written at manufacture or once thereafter.
5. Zero page — the first 256 bytes of memory on a 6502 — a short-addressed region used as extended register space.

14.12 The personal computer revolution#

PLAIN14.12.1 in simple words#

1. In December 1974 a magazine cover changed everything. The January 1975 issue of Popular Electronics showed a blue box called the Altair 8800.
2. It was made by MITS, a small company in Albuquerque run by Ed Roberts, and it cost $439 as a kit or $621 assembled.
3. It had an Intel 8080, 256 bytes of memory, no keyboard, no screen and no storage. You entered programs with switches and read answers as lights.
4. MITS hoped to sell a few hundred. They had 1,000 orders in February 1975 alone, and had shipped over 5,000 by August.
5. Two young men in Boston saw the cover. Bill Gates and Paul Allen phoned MITS and claimed to have a BASIC language for it. They did not.
6. They wrote it in eight weeks on a simulated 8080 running on a PDP-10, having never touched an Altair. Paul Allen flew to Albuquerque, and it worked first time.
7. They founded Microsoft on 4 April 1975 to sell it.
8. On 5 March 1975 about thirty hobbyists met in Gordon French’s garage in Menlo Park, California. That was the first Homebrew Computer Club.
9. Two members were Steve Wozniak and Steve Jobs. Apple came out of that room.
10. From there the story accelerates: Apple in 1976 and 1977, three rival home computers in 1977, the first spreadsheet in 1979, and IBM’s entry in 1981.

PLAIN14.12.2 a picture in your head#

1. Think of the first cheap cars. They were sold as parts, bought by mechanics, and driven by people who could fix them at the roadside.
2. Then somebody sold a car with a roof, a starter motor and a key, and ordinary people bought it.
3. The Altair is the kit. The Apple II with a plastic case and a power supply is the car with a roof.
4. But nobody buys a car because it has a roof. They buy it to get somewhere.
5. The place people wanted to get to turned out to be a spreadsheet.

Where this comparison breaks: cars had obvious uses before they were easy to drive. In 1977 nobody could say what a home computer was for. That question was genuinely open, and shop owners struggled to answer it, until VisiCalc gave a concrete answer for businesses in 1979.

PLAIN14.12.3 a worked example#

1. The three machines of 1977, often called the 1977 trinity, arrived within months of each other and defined the market.

2. All three included BASIC in read-only memory, so the machine was usable the moment it was switched on.
3. The TRS-80 sold about 55,000 in its first year, helped by Radio Shack’s 3,000 stores. Over 200,000 were sold in total.
4. The Apple II sold about 6 million across all models between 1977 and 1993.
5. Apple Computer had been founded on 1 April 1976 by Steve Jobs, Steve Wozniak and Ronald Wayne. Wayne sold his ten per cent share back for $800 twelve days later.
6. The Apple I went on sale in July 1976 at $666.66. About 200 were built. The first order was 50 units from Paul Terrell’s Byte Shop at $500 each, on condition they were assembled boards.

PLAIN14.12.4 what is really happening inside#

1. VisiCalc was released on 17 October 1979 for the Apple II, at under $100.
2. Dan Bricklin, a Harvard business student, imagined it. Bob Frankston wrote it. Personal Software, later VisiCorp, published it.
3. It is the first killer application: a program so useful that people buy the hardware in order to run it.
4. More than 700,000 copies sold in six years. Reports at the time suggested over a quarter of Apple IIs sold in 1979 were bought to run it.
5. That commercial signal is what brought IBM in. IBM normally took years to develop a product. This time it gave a team in Boca Raton, Florida, one year.
6. The project was code-named Chess, led by William C. Lowe and then Don Estridge. To hit the deadline they broke every IBM rule and bought parts from outside.
7. The IBM PC model 5150 was announced on 12 August 1981: an Intel 8088 at 4.77 megahertz, 16 kilobytes of memory as standard, from $1,565.
8. The decision that made history was openness. IBM published a technical reference manual containing the full circuit diagrams and the complete source listing of the BIOS, the start-up firmware.
9. That let anyone build add-in cards and software. It also let anyone build a copy of the machine, provided they could produce a BIOS without copying IBM’s code.
10. Compaq did exactly that. One team of engineers documented what the IBM BIOS did; a second team, who had never seen IBM’s code, wrote a new one from that description alone. That procedure is called clean room design.
11. The Compaq Portable was announced in November 1982 and shipped in March 1983 at $2,995. Compaq sold 53,000 units and made $111 million in its first year, an American record at the time.
12. Once Phoenix Technologies began selling a compatible BIOS to anyone in 1984, the clone industry exploded, and the PC stopped being IBM’s product and became an industry standard.

TECHNICAL14.12.5 the engineer’s version#

1. How Microsoft got MS-DOS, with dates, because this is constantly told wrong.
   1. IBM approached Digital Research, maker of CP/M, the dominant 8-bit operating system. No deal was reached. The famous story that Gary Kildall went flying instead of meeting IBM is disputed by participants and should not be stated as fact.
   2. Microsoft agreed to supply an operating system it did not have.
   3. Tim Paterson at Seattle Computer Products had written 86-DOS, nicknamed QDOS for Quick and Dirty Operating System, starting in April 1980 and shipping in August 1980. It was modelled on CP/M but for the 8086.
   4. In December 1980 Microsoft licensed it non-exclusively for $25,000.
   5. In July 1981 Microsoft bought all rights for a further $50,000, and hired Paterson.
   6. IBM shipped it in August 1981 as PC DOS 1.0.
2. The commercial masterstroke was not the purchase. It was the contract: Microsoft licensed to IBM non-exclusively and kept the right to sell MS-DOS to anyone else. Within a year it had licensed it to more than 70 companies.
3. So Bill Gates did not write MS-DOS, and Microsoft did not originally own it. Both facts are commonly got wrong.
4. IBM PC 5150 hardware, for the record: Intel 8088, 4.77 MHz, 8-bit external bus with 16-bit internal registers, 1 megabyte address space, five expansion slots, cassette port, and 16 to 64 kilobytes on the system board.
5. The 8088 was chosen over the faster 8086 precisely because its 8-bit external bus let IBM reuse cheap existing 8-bit support chips.
6. Three operating systems were offered at launch: PC DOS at $40, CP/M-86 at $240, and UCSD p-System. Price decided the outcome.
7. The legal foundation for clones is that copyright protects the expression of the BIOS code, not the interface it presents. Clean room procedure documents that the new expression was created independently. This principle was reinforced in Computer Associates v. Altai in 1992.

WORDS14.12.6 remember these#

1. Kit computer — a computer you solder together — an unassembled system such as the Altair 8800 of January 1975.
2. Killer application — the program worth buying a computer for — software whose demand drives hardware adoption, first VisiCalc in 1979.
3. BIOS — the start-up firmware in a PC — Basic Input/Output System, the low-level routines between hardware and operating system.
4. Clean room design — copying behaviour without copying code — reimplementation from a specification written by a separate team with no access to the original code.
5. Clone — a machine that runs the same software — a hardware-compatible system using an independently written BIOS.
6. Open architecture — published plans anyone can build for — documented interfaces and bus specifications permitting third-party hardware.

14.13 The graphical interface#

PLAIN14.13.1 in simple words#

1. Until the 1980s you talked to a computer by typing commands and reading text.
2. The alternative was demonstrated on 9 December 1968 in San Francisco, by Douglas Engelbart of the Stanford Research Institute.
3. In ninety minutes he showed a mouse, windows on screen, clickable linked text, shared editing of a document with a colleague miles away, and video of that colleague on the same screen.
4. His system was called NLS, the oN-Line System. The demonstration is now called the Mother of All Demos.
5. Almost nothing he showed was available to buy for fifteen years.
6. In 1970 Xerox opened a research centre in Palo Alto, called PARC, and hired many of Engelbart’s people.
7. On 1 March 1973 PARC had the Alto working: a personal computer with a screen made of individually controlled dots, a mouse, and a network socket.
8. PARC also produced Smalltalk, a programming system by Alan Kay’s group in which overlapping windows, pop-up menus and the whole desktop idea were worked out.
9. Steve Jobs visited PARC in December 1979 and saw all of it.
10. Apple shipped the Lisa on 19 January 1983 at $9,995, and the Macintosh on 24 January 1984 at $2,495.
11. Microsoft shipped Windows 1.0 on 20 November 1985 at $99, and Apple sued in 1988.

PLAIN14.13.2 a picture in your head#

1. A command line is like ordering food by writing a note in a language the kitchen understands, and getting a note back.
2. A graphical interface is like walking to a counter where the dishes are laid out, and pointing.
3. Pointing needs three things: a picture of the choices, a way to point, and instant feedback when you do.
4. The picture needs a screen where every dot is separately controlled, which is called a bitmap display. That is far more memory than a text screen.
5. The pointing needs a mouse. The feedback needs the machine to redraw fast enough that the arrow feels attached to your hand.
6. That combination is why the idea waited from 1968 to 1984. It needed memory to become cheap.

Where this comparison breaks: pointing at a counter shows you only what is on display. A command line can express things no menu contains, such as “rename every file recorded before March and move it”. Neither replaced the other, which is why every professional operating system in 2026 still has a terminal.

PLAIN14.13.3 a worked example#

1. Memory is the whole reason for the delay. Count the bits.
2. A 1970s text terminal: 80 columns by 24 rows of characters, one byte each. 80 times 24 is 1,920 bytes.
3. The Alto’s screen: 606 by 808 dots, one bit each. That is 489,648 bits, about 61 kilobytes.
4. So a graphical screen needs about 32 times the memory of a text screen just to hold the picture, before any program runs.
5. The original Macintosh screen is 512 by 342, one bit per dot, which is 21,888 bytes out of a total of 128 kilobytes of memory.
6. Seventeen per cent of the whole machine was the picture on the screen.

PLAIN14.13.4 what is really happening inside#

1. Engelbart and Bill English built the first mouse prototype in 1964, a wooden box with two wheels at right angles. The patent, number 3,541,541, was granted on 17 November 1970.
2. Xerox did sell a graphical machine. The Xerox Star, formally the 8010 Information System, launched in April 1981 with icons, windows, folders and a two-button mouse.
3. It cost $16,595 per workstation and needed a network of them to be useful. It sold poorly. Xerox invented the future and failed to price it.
4. The Apple visit of December 1979 was not theft. Xerox was given the right to buy 100,000 Apple shares before Apple’s public offering, in exchange for showing its work.
5. Apple’s engineers then built things PARC had not: the pull-down menu bar fixed at the top, click-and-drag direct manipulation, overlapping windows on a cheap machine, and the trash can.
6. The Lisa failed commercially. It was slow, its own disk format was unreliable, and it cost as much as a house deposit. Estimates of units sold range from about 10,000 to 60,000.
7. The Macintosh succeeded, helped by the “1984” advertisement directed by Ridley Scott, shown once during the Super Bowl on 22 January 1984.
8. Microsoft’s Windows 1.0 could not overlap its windows. They were tiled, side by side, partly for technical reasons and partly because of Apple’s licensing terms. Overlapping arrived with Windows 2.0 in December 1987.

TECHNICAL14.13.5 the engineer’s version#

1. Alto specification: introduced 1 March 1973, designed largely by Charles P. Thacker; 128 kilobytes of memory expandable to 512; a 2.5 megabyte removable disk cartridge; a portrait bitmap display; Ethernet at 2.94 megabits per second, invented at PARC by Robert Metcalfe and David Boggs, whose founding memo is dated 22 May 1973. About 2,000 were built.
2. Smalltalk-72, Smalltalk-76 and Smalltalk-80 came from Alan Kay’s Learning Research Group. Smalltalk is where object-oriented programming, the model-view-controller pattern and the modern integrated development environment were worked out together.
3. Bravo, written by Charles Simonyi in 1974 for the Alto, is the first what-you-see-is-what-you-get document editor. Simonyi later joined Microsoft and led Word and Excel.
4. Windows release history, with dates:

5. Windows 95 sold about one million copies in four days and was supported by an advertising campaign costing hundreds of millions of dollars. It is the moment the graphical interface became the default for ordinary computers.
6. Apple Computer, Inc. v. Microsoft Corp., the look-and-feel case:
   1. Filed in 1988 against Microsoft and Hewlett-Packard.
   2. Apple listed 189 interface elements it said were copied.
   3. Microsoft’s defence rested on a 1985 agreement licensing Apple interface elements for Windows 1.0. The district court found 179 of the 189 elements were covered by that licence.
   4. For the rest, the court applied a “virtual identity” standard, because the remaining similarities were unprotectable ideas or the only practical way of expressing them.
   5. The Ninth Circuit affirmed in 1994. The Supreme Court declined to hear the case in 1995.
   6. Xerox sued Apple during the same period and lost, mainly on timing.
   7. All remaining disputes were settled in August 1997, with Microsoft investing $150 million in non-voting Apple stock.
7. The lasting legal principle: an interface idea, such as overlapping windows or icons for files, is not copyrightable. Its specific artwork and code are.

WORDS14.13.6 remember these#

1. Graphical user interface — pointing at pictures instead of typing — a GUI, using windows, icons, menus and a pointer.
2. Bitmap display — a screen where every dot is separately set — a raster framebuffer with one or more bits per pixel.
3. Mouse — the pointing box — an X-Y position indicator, prototyped by Engelbart and English in 1964, patented in 1970.
4. WYSIWYG — the screen looks like the printout — what you see is what you get, first in Bravo on the Alto in 1974.
5. Desktop metaphor — files as paper on a desk — the interface model of the Xerox Star, 1981, and its successors.
6. Look and feel — the overall feel of an interface — the contested subject of Apple v Microsoft, largely held unprotectable in 1994.

14.14 Networks, the web and mobile: the spine only#

PLAIN14.14.1 in simple words#

1. This section is deliberately short. Chapter 30 tells the story of networks and the web in full, and Chapter 31 tells the story of mobile.
2. Here you get only the spine, so the timeline in this chapter is not missing a limb.
3. Networks: in 1969 the United States funded ARPANET, a network joining four university computers so researchers could share machines.
4. The first message was sent on 29 October 1969, from UCLA to Stanford Research Institute, by a student named Charley Kline. He typed L, then O, and the system crashed before the G.
5. The web: in March 1989 Tim Berners-Lee, a British engineer at CERN in Switzerland, wrote a proposal for a linked document system. His boss’s note on it read “vague but exciting”.
6. Mobile: phones went from car-boot-sized analogue radios in 1983 to pocket computers with a screen in 2007.

PLAIN14.14.2 a picture in your head#

1. A network is a postal system that never asks who you are, only where the letter goes.
2. The web is a library laid on top of that postal system, where every book can contain the address of another book.
3. Mobile is the same postal system with the address being a moving person rather than a fixed building.
4. Each layer is built on the one below and does not know how it works.

Where this comparison breaks: the post office delivers whole letters, while a network chops your message into small numbered pieces that may travel by different routes and arrive out of order. Reassembly is the receiver’s job.

PLAIN14.14.3 a worked example#

1. You can watch this history with one command on your own machine.

traceroute github.com
 1  192.168.0.1
 2  172.31.0.17
 3  137.97.29.249
 ...
 7  ae66-0.del01-96cbe-1b.ntwk.msn.net
 8  be23.rwa02.bom01.ntwk.msn.net
10  be5.ibr02.pnq21.ntwk.msn.net

2. Every line is a machine that looked at your packet, decided where next, and passed it on.
3. That “look and forward” behaviour is exactly what the first ARPANET nodes did in 1969, running on Honeywell minicomputers called Interface Message Processors.
4. The addresses beginning 172 are private ranges reserved by RFC 1918. The names ending msn.net are one company’s naming convention, not a standard.
5. Fifty-seven years separate the first hop and the last, and the basic idea has not changed.

PLAIN14.14.4 what is really happening inside#

1. The spine of networking, in eight dates.
2. 1969: ARPANET’s first four nodes.
3. 1974: Vinton Cerf and Robert Kahn publish the design of TCP, the protocol that lets separate networks join into one internet.
4. 1 January 1983: the whole ARPANET switches to TCP/IP on a single day, still called the flag day.
5. 1983 and 1984: Paul Mockapetris designs the Domain Name System, so people can type names instead of numbers.
6. March 1989: Berners-Lee’s proposal at CERN. December 1990: the first web server and browser run.
7. 30 April 1993: CERN places the web software in the public domain, charging nothing, for ever. That decision is why the web is not a product.
8. 1993: the Mosaic browser from NCSA, written by Marc Andreessen and Eric Bina, puts images and text on one page and the public arrives.
9. 1998: Google is founded, and finding things becomes the main problem.

TECHNICAL14.14.5 the engineer’s version#

1. Mobile generations, with the years the first commercial networks opened:

2. The first GSM call was made in Finland in 1991. The first SMS text message was sent in December 1992 and read “Merry Christmas”.
3. 3G opened commercially with NTT DoCoMo’s FOMA service in Japan in October
   2001. The first commercial 4G LTE networks opened in Stockholm and Oslo in December 2009.
4. Two structural facts worth carrying forward. First, the internet has no centre and no owner; it is an agreement between networks to pass traffic. Second, the web is one application running on it, alongside email, video calls and file transfer.
5. Chapter 30 covers packets, routing, TCP, DNS, HTTP and TLS properly. Chapter 31 covers radio, cells, handover and the mobile stack.

WORDS14.14.6 remember these#

1. ARPANET — the 1969 research network — the packet-switched predecessor of the internet, funded by the US Advanced Research Projects Agency.
2. Packet — a small numbered piece of your message — a routed protocol data unit with header and payload.
3. TCP/IP — the agreement that joins networks together — the Transmission Control Protocol and Internet Protocol suite, universal from 1 January 1983.
4. DNS — the phone book turning names into numbers — the Domain Name System, designed by Paul Mockapetris in 1983 and 1984.
5. World Wide Web — linked pages on top of the internet — the hypertext system proposed by Berners-Lee at CERN in March 1989, released freely in 1993.

14.15 The last twenty years#

PLAIN14.15.1 in simple words#

1. For thirty years computers got faster in the simplest way: the clock ticked faster every year.
2. Around 2004 that stopped. Pushing the clock higher made chips too hot to cool. Intel cancelled its next high-clock designs in May 2004.
3. The answer was to put several complete processors, called cores, on one chip. Mainstream dual-core desktop chips arrived in May 2005.
4. This shifted the burden onto software. A single-threaded program gets no faster on a four-core chip.
5. On 9 January 2007 Apple announced the iPhone, which shipped on 29 June
   2007. In September 2008 the first Android phone was announced.
6. From 2006, Amazon began renting computing by the hour. Storage, called S3, opened on 14 March 2006, and rentable servers, called EC2, in August 2006. That is cloud computing.
7. Around 2011 the world began shipping more smartphones than personal computers, and never went back.
8. Graphics chips, built to draw triangles for games, turned out to be ideal for the arithmetic behind machine learning.
9. Since 2012 that has driven the largest hardware build-out in the industry’s history.
10. In 2026 the fastest computer on the public TOP500 list is a Chinese machine called LineShine, at about 2.2 exaflops.

PLAIN14.15.2 a picture in your head#

1. Imagine a restaurant kitchen with one chef who cooks faster every year.
2. Eventually the chef cannot move any faster without catching fire.
3. So you hire four chefs. The kitchen can now serve four tables at once.
4. But one complicated dish still takes one chef the same time. Four chefs cannot chop one onion four times as fast.
5. That is why your computer has eight or sixteen cores and some programs still feel exactly as slow as they did in 2010.

Where this comparison breaks: chefs share one kitchen. Cores share caches and one path to memory, so adding cores can even slow things down through contention. Splitting work also costs time in coordination, which is why Amdahl’s law, stated by Gene Amdahl in 1967, says the sequential part of a program sets a hard ceiling on any speed-up.

PLAIN14.15.3 a worked example#

1. Amdahl’s law in numbers. Suppose 90 per cent of a program can be split across cores and 10 per cent cannot.
2. Speed-up with N cores is 1 divided by (0.1 plus 0.9 over N).

3. With a tenth of the work stuck in sequence, a thousand cores buy you less than ten times the speed.
4. That table explains the whole software industry since 2005. It is also why machine learning suits GPUs so well: its core operation, multiplying large matrices, is nearly all parallel.

PLAIN14.15.4 what is really happening inside#

1. Why the clock stopped rising: a rule called Dennard scaling used to say that as transistors shrank, power per unit area stayed constant. It broke down around 2005 because leakage current stopped shrinking with size.
2. So power density rose with frequency, and the heat became unremovable with air cooling.
3. Clock speed comparison: a Pentium 4 reached 3.8 gigahertz in 2004. High-end desktop chips in 2026 boost to roughly 5 to 6 gigahertz. That is under twice as fast in twenty-two years.
4. Core counts in the same period went from 1 to 16 or 24 on a desktop, and to well over 100 on server parts.
5. Cloud computing changed who owns machines. Before 2006, needing 500 servers for one week meant buying 500 servers. After 2006 it meant a credit card and ten minutes.
6. Phones changed what a computer looks like. The processor in a modern phone uses the ARM instruction set, designed in Cambridge in 1985 for the Acorn Archimedes, and licensed rather than manufactured by its owner.
7. Apple moved its own laptops and desktops to ARM chips of its own design with the M1 in November 2020, ending nearly fifteen years of Intel Macs.

TECHNICAL14.15.5 the engineer’s version#

1. Multicore timeline: IBM POWER4, the first commercial dual-core general purpose processor, shipped in 2001. Mainstream x86 dual-core parts, the AMD Athlon 64 X2 and the Intel Pentium D, both launched in May 2005.
2. GPU computing timeline: Nvidia released the CUDA software development kit on 15 February 2007 for the G80 generation, starting with the GeForce 8800.
3. The result that changed the field was AlexNet, which won the ImageNet competition in 2012 by a large margin, trained on two consumer graphics cards. The 2017 paper Attention Is All You Need introduced the transformer architecture behind current language models.
4. Where things stand in 2026, separated honestly:
   1. Established fact: Nvidia’s Blackwell B100, from 2024, carries about 208 billion transistors. Apple’s M3 Ultra, from 2025, carries about 184 billion. TSMC’s 3-nanometre class process is in high-volume production and 2-nanometre class parts are entering it.
   2. Established fact: on the June 2026 TOP500 list the leading machine is LineShine at the National Supercomputing Centre in Shenzhen, China, at 2,198 petaflops, ahead of El Capitan at Lawrence Livermore at 1,809 and Frontier at Oak Ridge at 1,353.
   3. Active research: whether current model architectures continue to improve with more data and compute, and what the true limits of low precision arithmetic are for training.
   4. Active research: quantum computing. Real machines exist and real results are published, but no general commercial workload runs faster on one today.
   5. Marketing claim: process node names such as “3 nanometre” no longer describe any physical dimension on the chip. They are commercial labels. Compare vendors by transistor density, not by node name.
5. Process node names are the clearest example in the industry of a marketing convention dressed as a specification. Say so when you hear it.
6. Tools that let you observe your own machine’s place in this story:

lscpu                 # cores, threads, caches, flags
cat /proc/cpuinfo     # per-core model and speed
nvidia-smi            # GPU model, memory, utilization
sysctl -n machdep.cpu.brand_string   # macOS CPU name

WORDS14.15.6 remember these#

1. Core — a complete processor inside the chip — an independent execution unit with its own registers and level-1 caches.
2. Dennard scaling — smaller transistors used to mean cooler chips — the constant power-density rule that failed around 2005.
3. Amdahl’s law — the slow part sets the limit — the speed-up bound imposed by the non-parallel fraction of a program, stated in 1967.
4. Cloud computing — renting computers by the hour — on-demand elastic provisioning of remote compute and storage, from AWS S3 in March 2006.
5. GPU — the graphics chip now used for mathematics — a throughput-oriented many-core processor, made general-purpose by CUDA in 2007.
6. Exaflop — a billion billion sums a second — 10 to the power 18 floating point operations per second.

14.16 One master timeline, 1801 to 2026#

PLAIN14.16.1 in simple words#

1. A timeline is not a list of gadgets. It is a list of moments when something became possible that had not been possible the day before.
2. Read it once for the shape, not for the detail.
3. You will see three kinds of entry: an idea written down, a machine that worked, and a product that sold.
4. All three are needed. An idea nobody builds is philosophy. A machine nobody buys is a museum piece.
5. You will also see long gaps. Between Babbage in 1837 and the Z3 in 1941 there are 104 years in which almost nothing happened in general computing.

PLAIN14.16.2 a picture in your head#

1. Think of a relay race where the baton is dropped for a century and picked up by somebody who never met the first runner.
2. Babbage’s designs were largely forgotten. Turing did not build on them. Zuse had not heard of them.
3. The same idea was found independently three or four times.
4. That is normal in this history. Ideas arrive when the parts to build them become cheap, not when someone first thinks of them.

Where this comparison breaks: it was not a single race. Several teams ran in parallel in different countries, often in secret, and the winner was often decided by who was allowed to publish.

PLAIN14.16.3 a worked example#

1. Take one row and unpack it: 1948, Manchester Baby runs.
2. The idea came from Turing in 1936 and the design from the EDVAC report in 1945, so the idea is 12 years older than the machine.
3. The enabling part was the Williams tube, a memory made from a war-surplus radar display.
4. So the row hides three things: a mathematician’s paper, a wartime component, and an engineering team.
5. Every row in the table below hides a story of that shape.

PLAIN14.16.4 what is really happening inside#

1. Four forces recur throughout this timeline.
2. Cost per operation falls, roughly by a factor of ten every five to seven years across the whole period.
3. Physical size falls, from a room, to a cabinet, to a board, to a chip, to part of a chip.
4. Who is allowed to use it widens: governments, then corporations, then departments, then individuals, then everyone with a pocket.
5. What it is used for widens from arithmetic, to records, to text, to pictures, to communication, to everything.
6. If you can name those four trends you can place any unfamiliar machine in this story within a decade of its real date.

TECHNICAL14.16.5 the engineer’s version#

1. A caution on any timeline: dates in computing are slippery because announcement, first working unit, first delivery and general availability can be years apart.
2. Where they differ in this chapter, the body text says which one is meant.
3. Where a claim is contested, such as who first built a computer, the chapter says so rather than picking a winner.

WORDS14.16.6 remember these#

1. Announcement date — when it was told to the world — the marketing date, often the earliest and least meaningful.
2. First working unit — when it actually ran — the engineering milestone used for priority claims.
3. General availability — when anyone could buy one — the commercial date, the one that changes an industry.
4. Priority claim — the argument about who was first — a disputed assertion that depends entirely on the definition chosen.

14.98 Common wrong ideas#

1. Wrong: Babbage built the Analytical Engine. Right: he designed it from 1837 and built only fragments. Neither of his engines was completed in his lifetime; he died in 1871. The Science Museum in London built Difference Engine No. 2 from his drawings, finishing the calculating section in 1991.
2. Wrong: the first computer bug was a moth. Right: the moth found in a Harvard Mark II relay on 9 September 1947 was taped into a logbook with the note “First actual case of bug being found”. The joke only works because engineers already said “bug” for a fault. Thomas Edison used the word that way in 1878.
3. Wrong: Grace Hopper coined the word bug. Right: she popularized the moth story. She did lead the team that built the first compiler, the A-0 system in 1952, and shaped COBOL from 1959, which is achievement enough.
4. Wrong: Bill Gates wrote MS-DOS. Right: Tim Paterson at Seattle Computer Products wrote 86-DOS in 1980. Microsoft licensed it for $25,000 in December 1980 and bought it outright for $50,000 in July 1981. Gates and Allen did write Altair BASIC themselves in 1975.
5. Wrong: Apple invented the graphical interface. Right: Douglas Engelbart demonstrated the core ideas on 9 December 1968 and Xerox PARC built them into the Alto in 1973 and sold the Star in 1981. Apple licensed, refined and priced it for ordinary buyers with the Lisa in 1983 and the Macintosh in
6. Wrong: Apple stole the interface from Xerox. Right: the December 1979 visit was arranged, and Xerox received the right to buy Apple shares in exchange. Xerox later sued Apple over the interface and lost, largely on timing.
7. Wrong: Al Gore said he invented the internet. Right: on 9 March 1999, on CNN, he said “I took the initiative in creating the Internet”, meaning legislation. Vinton Cerf and Robert Kahn, who designed TCP/IP, publicly defended his record. His 1991 High Performance Computing Act helped fund the Mosaic browser.
8. Wrong: ENIAC was the first computer. Right: that depends entirely on the definition. The Z3 of 1941 was program-driven, the Atanasoff-Berry machine of 1942 was electronic, Colossus of 1943 was electronic and digital, and the Manchester Baby of 1948 was the first to run a stored program. ENIAC was the first general-purpose electronic machine put to broad use.
9. Wrong: computers have always used binary. Right: ENIAC, the Harvard Mark I, the IBM 650 and many other early machines were decimal. Binary won because it halves the components needed per digit and matches switching circuits.
10. Wrong: Silicon Valley is named after a company. Right: the journalist Don Hoefler used the phrase in Electronic News on 11 January 1971. Silicon is the element the chips are made from, and the valley is the Santa Clara Valley in California.

14.99 Chapter summary in 20 lines#

1. Counting aids came first: tally sticks, the abacus, and the geared Antikythera mechanism of roughly 100 BCE, found in a wreck in 1901.
2. Napier’s bones of 1617, the slide rule of about 1622 and Leibniz’s 1703 paper on binary set the mathematical stage.
3. Schickard in 1623, Pascal in 1642 and Leibniz in 1673 built calculators that worked but could not be reprogrammed.
4. Jacquard’s loom head, patented in 1804, separated the machine from its instructions. That is the founding idea of the whole field.
5. Babbage designed the Difference Engine from 1822 and the Analytical Engine from 1837, with mill, store, card reader and printer. Neither was finished.
6. Ada Lovelace’s 1843 notes contain the first published algorithm and the first clear statement that a computer could handle symbols, not just quantities.
7. Hollerith’s punched card machines processed the 1890 United States census. His company merged in 1911 and was renamed IBM in 1924.
8. Gödel showed in 1931 that proof has limits. Church and Turing showed in 1936 that computing has limits too, and Turing’s model described every later machine.
9. Turing complete means able to simulate a universal Turing machine, and the bar is so low that card games and spreadsheet formulas clear it by accident.
10. The war produced the Bombe in 1940, Zuse’s Z3 in 1941, the Atanasoff-Berry machine in 1942, Colossus in 1943 and 1944, and the Harvard Mark I in 1944.
11. ENIAC, shown in 1946, was electronic and general-purpose but was programmed by rewiring, work done by six women who went uncredited for forty years.
12. The 1945 First Draft of a Report on the EDVAC set out the stored-program design that everything since has followed, under a disputed name.
13. The Manchester Baby ran the first stored program on 21 June 1948, EDSAC gave the first service in 1949, and the Ferranti Mark 1 was sold in 1951.
14. UNIVAC I predicted the 1952 election correctly and was overruled live on air, which is how the public learned what a computer was.
15. The transistor of 1947 and IBM’s System/360 of 1964 turned computing into an industry with compatible product families and portable software.
16. Kilby in 1958 and Noyce in 1959 put whole circuits on one chip, and Hoerni’s planar process made them manufacturable.
17. The Intel 4004 of 1971 put a processor on one chip; the $25 MOS 6502 of 1975 made home computers affordable.
18. The Altair of 1975, Apple II of 1977, VisiCalc of 1979 and the IBM PC of 1981 with its published design created the personal computer industry.
19. Engelbart in 1968 and Xerox PARC in 1973 invented the graphical interface; Apple sold it in 1984 and Windows 95 made it universal in 1995.
20. Since 2005 progress has come from many cores, rented cloud machines, phones and graphics chips repurposed for machine learning, with the June 2026 TOP500 led by LineShine at about 2.2 exaflops.