Before the Transistor

Open the door of your microwave oven and look at the back wall. Behind it sits a vacuum tube, the same family of device that ran every radio, television, and computer on Earth until the late 1950s. It is called a magnetron, it puts out about a kilowatt of radio energy to heat your dinner, and no transistor made today can do that job as cheaply. Now go and start your car. The click you hear when you turn the key is a relay, and so are the switches behind the headlamps, the horn, and the fuel pump: a device older than the vacuum tube, older even than the light bulb, still doing the job because nothing has beaten it at that job.

I have wanted to write this series for a long time. Jon Stokes's book Inside the Machine is the thing that taught me how a processor actually works. It starts with nothing more than a light switch and ends with the real chips of its era, the Pentium 4 and the PowerPC G5, and by the end you can read a spec sheet and understand what every line means. It is also, by design, a book about the CPU. The GPU gets a passing mention. This series is my attempt at the same journey for the graphics processor, from the switch all the way up to Nvidia's Ada, Hopper, Blackwell, and Vera Rubin. No prior electronics needed. If you have read ML Basics or LLM Basics, you already know the one calculation we will follow the whole way: the neuron's w × x + b, a multiply and an add. A language model does trillions of those per second. The question this series answers is how a slab of silicon manages it.

Stokes starts with the switch. I want to start one step earlier, with the switches that came before the transistor, because the transistor only makes sense as an answer to their problems, and because those older devices are still all around you, doing the jobs where they were never beaten.

The one thing electronics needs

Strip away the detail and every electronic device, from a hearing aid to a data centre, is built on a single trick: a small electrical signal controlling a large one. A radio takes a signal from the aerial too faint to hear and makes it strong enough to drive a loudspeaker. A telephone repeater takes a voice that has faded over a hundred miles of wire and restores it. A computer takes the tiny voltage that means "1" on one wire and uses it to decide whether another wire will be "1" or "0". The first two are called amplification, the third is switching, and they are the same trick at different settings. A component that can do it is called active, and the entire history of electronics is the history of three of them: the relay, the vacuum tube, and the transistor.

The relay: electricity controlling electricity

The oldest is the relay, invented in the 1830s for the telegraph. It is an electromagnet next to a switch. Run a current through the coil and the magnet pulls a metal arm across, closing the contacts. Stop the current and a spring pulls the arm back. A small current in the coil controls a large current through the contacts, and crucially, the two circuits never touch. Joseph Henry is usually credited with the first, around 1835, built so that a faint telegraph signal arriving after miles of wire could close a fresh circuit with a fresh battery and send itself onward. That is where the name comes from: it relays the message.

The relay's virtues are still its virtues today. It will switch enormous currents, hundreds of amps if you make it big enough, from a signal of a few milliamps. It isolates the controlling circuit from the controlled one completely, so a delicate electronic board can switch mains power without the mains ever reaching it. It is indifferent to voltage spikes, heat, and abuse that would destroy a semiconductor. It is cheap. And when it is off, it is off: a physical air gap, not a component that merely conducts very little. This is why your car's starter motor, which draws a couple of hundred amps, is switched by a relay that the ignition key controls with a trickle. It is why the compressor in an air conditioner starts with an audible clunk, why an old thermostat clicks, why a washing machine's programme drum used to sound like a clock, and why the boxes at an electrical substation that decide whether to cut off a faulty line are still called protective relays, even now that most of them are computers.

Its limits are just as physical. A relay has to move a lump of metal, which takes a few milliseconds, so it can switch perhaps a few hundred times a second. The contacts wear, pit, and eventually weld, so it has a lifetime measured in millions of operations rather than forever. It is the size of a sugar cube at best. It draws a steady current to hold the arm in place. And it cannot amplify a continuously varying signal at all: it is open or closed, nothing in between, so it is useless for radio or telephones.

People did build computers out of them. Konrad Zuse's Z3 in 1941, the first working programmable computer, used about 2,600 relays and could manage a multiplication in three seconds. The Harvard Mark I in 1944 ran on 3,500 relays and could be heard from the next room. And the most famous computer bug was literally a moth, crushed in a relay of the Mark II in 1947 and taped into the logbook as the "first actual case of bug being found", which was a joke, because engineers had been calling faults bugs since Edison. A relay computer is a wonderful thing to listen to and a hopeless thing to scale: at a few hundred operations a second, our running calculation w × x + b would take most of a second, and a modern GPU does tens of trillions of them in that time.

The vacuum tube: the first amplifier

The device that made electronics possible was found by accident. Around 1883 Thomas Edison noticed that a current would flow across the vacuum inside a light bulb from the glowing filament to a separate metal plate, but only in one direction. He patented it and did nothing with it. In 1904 John Ambrose Fleming used the effect to build a diode, a one-way valve for electricity, which is why British engineers call these devices valves to this day. Then in 1906 Lee de Forest put a third element between the filament and the plate, a wire mesh he called the grid, and found that a small voltage on the grid controlled a large current from filament to plate. That device, the triode, is the ancestor of every transistor in your graphics card.

Here is how it works, because the mechanism carries straight over. Heat a metal wire, the cathode, until it glows, and it boils off electrons into the surrounding vacuum. Put a positively charged metal plate, the anode, nearby and the electrons stream across to it: a current. Now put the grid in the path. Make the grid slightly negative and it repels electrons, choking the stream. Make it less negative and the stream flows freely. The grid draws almost no current itself, so a tiny signal on it, a faint radio wave or a whisper from a microphone, is reproduced as a large swing in the plate current. That is amplification, and it can also be switching: drive the grid hard negative and the tube is off, drive it to zero and the tube is on.

The garden tap is the analogy that carries this whole series. The handle is the grid, the water is the electron stream, and turning the handle moves no water itself, it just opens or closes the path. Hold onto that picture, because Part 2 uses exactly the same tap to explain the transistor, and the point of the whole story is that the tap stayed the same while everything around it changed.

What the tube made possible

The triode changed the world faster than any component before or since. In 1915 it let a telephone call cross the American continent for the first time, with three tube repeaters along the way, at Pittsburgh, Omaha, and Salt Lake City, restoring a voice that thousands of miles of copper had worn down to almost nothing. From 1920 it made radio broadcasting possible, both the transmitters and the sets in living rooms. It made television, radar, the electric guitar, sound in cinemas, and long-distance telephony. And because it could switch in a microsecond where a relay took a millisecond, it made the electronic computer.

Colossus, the British code-breaking machine of 1943, used 1,600 tubes in its first version and 2,400 in its second. ENIAC, completed in 1945 in Pennsylvania, used 17,468 of them, could do 5,000 additions a second, filled a room of about 170 square metres, and drew 150 kilowatts. Nearly every computer of the 1950s was a tube computer: UNIVAC, the IBM 700 series, the early Manchester machines. The largest of them all was the SAGE air defence system, which went into service in 1958 with around 50,000 tubes per computer, two computers per site, and a power bill of three megawatts for the site. It was the biggest computer ever built, by most measures it still is, and it was obsolete before it was finished, for reasons the next section explains.

The tube also never went away, and this is the part people miss. It survived wherever the job was too hot, too powerful, or too high in frequency for a semiconductor to do cheaply. The magnetron in your microwave is a tube. The X-ray tube in a hospital CT scanner, a dentist's surgery, and an airport baggage scanner is a tube, pumping electrons into a metal target at a hundred kilovolts. Communications satellites amplify their downlink with travelling-wave tubes, because for decades nothing else could produce that power at those frequencies with that efficiency in orbit, and many still do. Broadcast radio and television transmitters ran on tubes the size of dustbins into the 2000s. Particle accelerators drive their beams with klystrons, tubes the size of a person. The fluorescent display on an older microwave clock or car dashboard is a small vacuum tube. And a large share of guitar amplifiers, and a smaller share of expensive audio equipment, still use tubes on purpose, because the way a tube distorts when overdriven is the sound of rock music, and no transistor circuit quite copies it. If you own a Fender or a Marshall, you own a handful of triodes that would be recognisable to de Forest.

What was wrong with it

Everything that was wrong with the tube followed from one fact: to boil electrons off the cathode, you have to heat it. A typical small tube burns a couple of watts in its heater alone before it does any work, and the largest single item in ENIAC's 150 kilowatts was the heaters, with a ventilation system to match. A tube takes half a minute to warm up before it works at all, which is why old radios and televisions came on slowly. And the heater, like the filament in a light bulb, eventually burns out. A good tube lasts a few thousand hours. That is fine for a radio with five tubes. It is a catastrophe for a computer with seventeen thousand, because with that many, one of them is always about to fail. ENIAC's engineers ran the heaters below their rated voltage and never switched the machine off, and still lost a tube every couple of days, each failure meaning a hunt through the racks. SAGE used tubes specially built for reliability and still needed a full-time crew replacing them.

Then there is size. A tube is a glass bottle with a vacuum in it, and a vacuum needs a certain volume to hold electrodes apart. The smallest practical tubes were the size of a thumb, and no amount of ingenuity made them the size of a grain of rice. A computer's power is set by how many switches it has, and with tubes, every switch was a thumb-sized, two-watt, glass object with a lifetime of a few years and a warm-up time of thirty seconds. You could build ENIAC. You could not build anything a hundred times bigger, and by the late 1950s engineers had a name for the wall they had hit. Jack Morton of Bell Labs called it the tyranny of numbers in 1958. Every extra switch added cost, heat, failure rate, and thousands of hand-soldered joints, and the joints failed too.

Try it below. Pick a technology and a number of switches, and see what the machine would cost you in power, space, and reliability. Then slide up to the 76 billion switches of the graphics card that this series is about.

The figures are rough, the trend is not. Tubes and relays are fine up to a few thousand switches, marginal at tens of thousands, and impossible beyond. Yet everything interesting a computer can do, including everything in this series, needs millions of switches at the very least. The tube got computing started and then stood in its way.

How the transistor fixed it

In December 1947 at Bell Labs, John Bardeen and Walter Brattain pressed two gold contacts onto a sliver of germanium and found that a current into one of them controlled a larger current through the other. Their boss, William Shockley, worked out the theory and a more practical design within months. The three shared the Nobel Prize for it. The device was named the transistor, and what it did was nothing new: a small signal controlling a large one, the triode's job exactly. What was new was what it did not need.

It did not need a heater, because the electrons were already free inside the semiconductor. That removed the warm-up time, most of the power, and most of the heat at a stroke. Early transistors used a few milliwatts against a tube's few watts, a thousandfold difference before anyone had tried to optimise anything. It did not need a vacuum or a glass envelope, so it was rugged and could be made small: the first commercial transistors were the size of a pea, and, unlike the tube, there was no physical reason they could not be made smaller still. And it had no filament to burn out. A transistor that is kept within its ratings does not wear out at all, which turned the tyranny of numbers on its head. A machine with a million transistors is not a machine with a million things waiting to fail.

The transistor had its own problems, and for a decade the tube kept its jobs. Early germanium transistors were expensive, around eighteen dollars each in 1950 against seventy-five cents for a tube, and their behaviour drifted with temperature. They could not handle high power or high frequencies, which is why transmitters, radar, and anything with kilowatts stayed with tubes, and why the microwave oven still does. And they were noisy in the electrical sense. So the transistor won its first markets where small and low-power mattered more than anything else: hearing aids, which people had been carrying around with a tube amplifier and a battery pack the size of a book, took their first transistor in 1952 and had gone almost entirely transistor by 1954. That year also brought the Regency TR-1, the first commercially made transistor radio, with four transistors, which sold for about fifty dollars and made the word "transistor" mean "radio" for a generation. The first transistor computer ran at Manchester University in 1953, Bell Labs' TRADIC followed in 1954 with 684 transistors (and one tube, to generate its clock), and by 1959 IBM's 7090 mainframe was transistorised. No serious computer used tubes again.

Two further steps made the transistor into the switch this series is about. In 1958 Jack Kilby at Texas Instruments put several components on a single sliver of germanium, still wired together by hand, and in 1959 Robert Noyce at Fairchild did it on silicon with the connecting wires printed on in the same process as the transistors. That is the integrated circuit, it was the real answer to the tyranny of numbers, because the soldered joints disappeared along with the tubes, and it is the subject of Part 2. And in 1959, at Bell Labs, Mohamed Atalla and Dawon Kahng built a transistor that worked by a different mechanism from Bardeen and Brattain's, controlling the current with an electric field across a thin insulator rather than with a current. That is the MOSFET, it draws almost nothing to hold its state, it shrinks better than anything else ever invented, and there are 76 billion of them in the chip on the cover of this series.

Where you will find each one today

The pattern in the last row is the lesson of this part. Each technology kept the jobs where its particular strength mattered more than its weaknesses. The relay kept isolation and brute current. The tube kept extreme power and frequency. The transistor took everything where the count mattered, and it turned out that in computing, the count is the only thing that matters.

Where this leaves us

Electronics needed a component that lets a small signal control a large one. The relay did it with a magnet and a moving arm, and could switch a starter motor but not a radio signal. The tube did it with a heated cathode in a vacuum, could amplify anything, and built the first computers, but every one was a hot, thumb-sized, glass object with a lifetime of a few years, and that put a ceiling of tens of thousands on how many switches a machine could have. The transistor did the same job in a solid, with no heater, no vacuum, and no wearing out, and the ceiling disappeared. The tube and the relay are still around us, doing the jobs where they were never beaten. But the machine this series is about could only be built from the third.

So the next part starts where Stokes did, with the switch itself: what a MOSFET is, why it can only be on or off, how it got from one to 76 billion, and what it costs, in heat, to flip it a few billion times a second.

Next: The Switch: the transistor as a switch, why computers count in twos, what "4 nanometres" does and does not mean, and the end of the free lunch that made GPUs inevitable.