Stardust Log

A record of stardust, quietly accumulating.
Timeline

Semiconductor Timeline, 1847–2035: From Boolean Logic to EUV

From turning logic into algebra in 1847 to wiring at a width of a few dozen atoms

Semiconductor timelines usually begin with the transistor in 1947. But a transistor could be used for arithmetic only because, a decade earlier, it had been shown mathematically that opening and closing a switch in a circuit is itself a logical operation. This timeline runs the mathematics that guarantees computation along the top of the axis and the devices that carry it along the bottom — 43 entries, from Boole turning logic into algebra in 1847 to wiring at a width of a few dozen atoms today.

Updated 2026-10-15/43 entries

TheoryDiscoveryMilestoneDeployment

Dashed borders and hollow dots mark projected, not-yet-achieved items. “Unverified” marks entries not yet confirmed against a primary source.

Read vertically View horizontally
1847–1931Two Streams, Begun Separately
1847
Theory

Boole makes true-and-false something you can calculate with

George Boole rewrote true-and-false reasoning so it could be manipulated with symbols, the way numbers are added and multiplied. He sketched the framework in The Mathematical Analysis of Logic (1847) and built it into a system in An Investigation of the Laws of Thought (1854). At the time this was pure logic; nobody was thinking about machines. Ninety more years would pass before the algebra was connected to an electrical circuit.

  • Until it met a circuit ~90 years

People and institutions: George Boole

1874
Discovery

A material that lets current through in only one direction

Ferdinand Braun reported that when a metal point touched a crystal such as galena (lead sulfide), current flowed in one direction but not the other. Conductivity that depends on direction made no sense under the electrical theory of the day. Nobody could explain it, yet the effect was put to work as a detector in radio receivers long before anyone understood it — a very semiconductor-like beginning.

  • Until it was explained 57 years

People and institutions: Ferdinand Braun

NoteThe original paper is F. Braun, Ueber die Stromleitung durch Schwefelmetalle, Ann. Phys. Chem. 153, 556–563 (1874). The publisher record could not be reached, so the IEEE Milestone proposal and the Computer History Museum entry are cited instead.

1906
Milestone

The triode — the first device that could both amplify and switch

Lee de Forest added a mesh electrode inside a vacuum tube. A small voltage on that grid controls a much larger current through the tube, which makes the device both an amplifier and an on-off switch. The first electronic computers would be built from these. But a vacuum tube ran hot, burned out, and was about the size of a light bulb.

People and institutions: Lee de Forest / John Ambrose Fleming

NotePatent US879532 (Space Telegraphy) was filed in 1907 and granted 18 February 1908; the Audion itself was announced in 1906.

1913
Theory

Proof that a single operation is enough to build all of logic

Henry Sheffer showed that one operation — "A and B are not both true" — is enough, when repeated and combined, to express everything in Boole's algebra. It is the operation later called NAND. This entry is usually left out of semiconductor histories, but it is decisive: for a chip designer it is the guarantee that laying down one kind of part, over and over, is sufficient to build any computer at all.

  • Operations required just one

People and institutions: Henry M. Sheffer

NoteThe same operation already appears in Charles S. Peirce's manuscript "A Boolian Algebra with One Constant" of around 1880. That manuscript stayed unpublished until his Collected Papers in 1933, by which time Sheffer's 1913 paper was the one the field had built on.

1918
Milestone

A circuit that can remember which state it is in

Eccles and Jordan wired two vacuum tubes so that each held the other down. As long as the power stayed on, the circuit rested in one of two states, and a pulse from outside flipped it to the other. Computation needs somewhere to keep intermediate results; this circuit, later called the flip-flop, became the smallest such place.

  • Information held 1 bit

People and institutions: William Eccles / Frank Wilfred Jordan

NoteSources differ on the date: 1918 (The Electrician) or 1919 (The Radio Review 1, 143–146). The year the circuit was devised, 1918, is used here.

1926
Theory

A patent for controlling current with a voltage, in a solid — but nothing was built

Julius Lilienfeld patented a device that would control current through a thin semiconductor film by applying a voltage from outside, with no vacuum tube involved. The idea is close to today's transistor. But neither the material purity nor the thin-film processing existed. A working version of this structure did not appear until 1960 — a gap of 34 years. An idea alone changes nothing if the materials cannot follow.

  • Until one was actually built 34 years

People and institutions: Julius Edgar Lilienfeld

NotePatent US1745175 was filed 8 October 1926 and granted 28 January 1930. There is no record that Lilienfeld ever built a working device.

1931
Theory

Quantum mechanics finally explains what a semiconductor is

Alan Wilson explained semiconductor behaviour in terms of allowed bands of electron energy separated by a forbidden gap. A semiconductor is not simply something halfway between a metal and an insulator: its gap is moderate, so adding a trace of impurity changes its conductivity enormously. Fifty-seven years after Braun's observation, the material became something you could design with rather than merely use.

People and institutions: Alan Herries Wilson

Read more on this site

The Rise of Quantum Mechanics, Part 4Free to Move, Yet No Current — Bloch, Wilson and the Electrons in Solids, 1928–1931

From Bloch's theory of 1928 to Wilson's account of insulators and semiconductors in 1931 — and how much less certain the path was than the usual summary suggests (silicon was then classed as a metal).

1936–1948Where Circuits Met Logic
1936Nov 1936
Theory

A definition of what it means to compute

Alan Turing imagined a very simple abstract machine that only writes and erases symbols on a paper tape, and argued that anything a person can compute by following steps, this machine can reproduce. He then showed that a single machine can imitate any of the others — a universal machine. That is the guarantee that you do not need to build a different machine for every calculation, and it is why one computer today can do everything.

  • Machines required one

People and institutions: Alan Turing

Connections to other timelines

Shares a basis withArtificial Intelligence1936What does it mean to compute? — defining one machine that can compute anything

The semiconductor and AI timelines both list the same 1936 paper by Turing. The semiconductor timeline takes it as the guarantee that a single computer can do everything; the AI timeline takes it as the starting point of the premise that, if intelligence is a kind of computation, a machine can in principle do it.

1936Dec 1936
Theory

Akira Nakashima, at NEC, treats relay circuits as algebra

Akira Nakashima, who designed telephone switching equipment, published with his colleague Masao Hanzawa a theory treating networks of relays — switches opened and closed by electromagnets — as symbolic operations, in the December 1936 Journal of the Institute of Telegraph and Telephone Engineers of Japan. The point was to design circuits by calculation rather than by intuition. It came roughly two years before Shannon's AIEE paper, but it was written in Japanese and barely reached the outside world.

  • Ahead of Shannon's AIEE paper by ~2 years

People and institutions: Akira Nakashima (中嶋章) / Masao Hanzawa (榛澤正夫)

NoteWhich came first depends on what is compared: Shannon's thesis is 1937 and his AIEE paper 1938, so Nakashima's December 1936 paper precedes both. But it was written in Japanese and was scarcely known outside Japan at the time. In the USSR, Viktor Shestakov had arrived at an equivalent algebra of relay circuits by 1935, but his work stayed unpublished until 1941, so it did not reach the field either.

1937
Theory

Switching a circuit on and off is itself a logical operation

In his master's thesis Claude Shannon showed that putting switches in series corresponds exactly to "A and B", and putting them in parallel to "A or B". The algebra of logic Boole had built ninety years earlier and an electrical circuit obey the same rules. Circuit design stopped being a matter of staring at diagrams and became a matter of manipulating expressions. On this timeline, this is the point where the upper band (mathematics) and the lower band (devices) finally cross.

  • Years after Boole 90

People and institutions: Claude Shannon

Note1937 is the year of the MIT master's thesis; the paper appeared as Trans. AIEE 57, 713–723 (1938).

1940
Discovery

A crack in a silicon rod reveals the p-n junction

At Bell Labs, Russell Ohl noticed that a cracked silicon rod produced a voltage when light fell on it. One side of the rod turned out to have a surplus of electrons (n-type) and the other a shortage (p-type), and the boundary between them was doing the work. That boundary — the p-n junction — is the mechanism behind one-way conduction. Every diode, transistor and solar cell since has been built on it.

People and institutions: Russell Ohl

NoteThe discovery dates to 1940; patent US2402662 was filed in 1941 and granted 25 June 1946.

1945Jun 1945
Theory

Put the instructions in the same place as the numbers

The draft John von Neumann circulated described a machine that keeps its instructions in memory rather than in its wiring, and reads and writes them exactly as it does data. Until then, giving a computer a different job meant rewiring it. Once the procedure lives in memory, the hardware stays put and only the work changes. It is Turing's universal machine rendered as something you could actually build.

People and institutions: John von Neumann

1946Feb 1946
Milestone

ENIAC — and the ceiling of what vacuum tubes could do

Unveiled at the University of Pennsylvania, ENIAC used about 17,468 vacuum tubes, weighed roughly 27 tonnes and filled a room. Ballistics calculations that took hours by hand took it tens of seconds. The limit was just as clear: tubes run hot and burn out, so adding more of them buys speed and failures in equal measure. The demand for a switch that was smaller and did not break became concrete here.

  • Vacuum tubes ~17,468
  • Weight ~27 t

People and institutions: John Mauchly / J. Presper Eckert

1947Dec 1947
Milestone

The point-contact transistor — a vacuum tube's job, done in a solid

At Bell Labs, John Bardeen and Walter Brattain pressed two gold contacts extremely close together onto a germanium surface and showed that a small current through one controlled a much larger current through the other. No vacuum, no heated filament. But point-contact devices were hard to reproduce and unsuited to manufacturing. This form never became the product; another idea was needed almost immediately.

People and institutions: John Bardeen / Walter Brattain / William Shockley

NoteThe device worked on 16 December 1947 and was demonstrated on 23 December; the paper appeared in Phys. Rev. 74, 230 (15 July 1948).

1948Jan 1948
Theory

The junction transistor — working inside the crystal, not on its surface

Seeing that point-contact devices were at the mercy of surface conditions, William Shockley proposed building the p-n junctions into the interior of the crystal instead: three stacked layers of p- and n-type material, with a small current in the middle layer controlling the whole. He published the theory in 1949; a working device followed in 1951. Because it does not depend on the accidents of a surface, this became the basic form of every mass-produced transistor since.

  • From concept to working device 3 years

People and institutions: William Shockley

NoteConceived January 1948; the theory was published in Bell Syst. Tech. J. 28, 435–489 (1949); a working device followed in 1951.

1948Jul 1948
Theory

Information gets a unit — and errors are shown to be correctable

Shannon showed that information can be measured in bits, and proved that even over a noisy channel, adding redundancy allows errors to be corrected reliably up to a definite limit. This matters for semiconductors too. Shrinking a device makes its behaviour less certain; it is the existence of error correction that makes shrinking that far a defensible decision rather than a reckless one.

  • Unit introduced the bit

People and institutions: Claude Shannon

Connections to other timelines

Shares a basis withArtificial Intelligence1948Measuring information with numbers — and producing English-like text from probabilities alone

The semiconductor and AI timelines both list the same 1948 paper by Shannon, but read different parts of it. The semiconductor timeline takes the proof that errors can be corrected even over a noisy channel; the AI timeline takes the definition of information as a quantity of probability, and the experiment that strings together English-like text from probabilities alone.

1952–1971Onto a Single Slab
1952May 1952
Theory

A prediction: "a solid block with no connecting wires"

Geoffrey Dummer of Britain's Royal Radar Establishment is reported to have said that electronic circuits would eventually be made on a single solid block, rather than as parts joined by wires. This was just after the transistor had begun to displace the vacuum tube. Dummer never got a working version built; the real thing appeared six years later.

People and institutions: Geoffrey W. A. Dummer

UnverifiedDummer is said to have predicted at a May 1952 components conference in Washington, D.C. that electronic circuits would be made on a solid block with no connecting wires. The proceedings could not be verified against a primary source, so no citation is given. The entry could be dropped without harming the timeline, but the fact that the integrated circuit was imagined six years before it was built is worth keeping.

1954May 1954
Deployment

From germanium to silicon — a device that tolerates heat

Gordon Teal of Texas Instruments announced a transistor made of silicon. Germanium becomes unusable around 75 °C; silicon keeps working close to 150 °C, and for military and automotive equipment that difference decided the matter. More important in the long run: silicon grows an oxide layer — essentially a glass skin — on its own surface, and every later manufacturing technique would be built on that.

  • Usable temperature vs germanium ~75°C → ~150°C

People and institutions: Gordon Teal / Texas Instruments

1954Jun 1954
Milestone

Diffusion — controlling how thick a layer is

Bell Labs established a way to drive gaseous impurities into a silicon surface inside a high-temperature furnace: diffusion. Until then, layers were formed by switching the melt while the crystal was being grown, and their thickness was largely a matter of luck. With diffusion, thickness is set by temperature and time. Since a transistor's speed depends on how thin its middle layer is, this meant performance had become something you could design.

People and institutions: Calvin Fuller / Bell Telephone Laboratories

1955
Milestone

Masking with oxide — making things only where you want them

Heat silicon and a film of silicon dioxide forms on its surface. Bell Labs found that this film blocks impurities, and that opening a window in it lets impurities in only underneath that window. What gets built, and where, could now be specified as a pattern. Combined with printing fine patterns photographically, this is the basis of all semiconductor manufacturing for the seventy years since.

People and institutions: Carl Frosch / Lincoln Derick / Bell Telephone Laboratories

1958Sep 1958
Milestone

Put the parts and the wiring on the same piece

Jack Kilby reasoned that transistors, resistors and capacitors could all be made on a single chip of semiconductor — and built one that worked. Making parts separately and joining them with wire means that the more parts you have, the more joints there are to fail. On one piece, the joints largely disappear. Kilby's prototype still had gold wires laid across it by hand, so it was not yet something you could manufacture.

  • Components in the prototype a handful

People and institutions: Jack Kilby / Texas Instruments

NoteFirst demonstrated 12 September 1958; patent US3138743 was filed February 1959 and granted 23 June 1964.

1959May 1959
Milestone

The planar process — keeping the surface flat

Rather than etching down to expose a junction, Jean Hoerni formed junctions beneath a flat surface kept sealed under its oxide. Junctions never meet the air, so their characteristics stay stable. And a flat surface means metal can be evaporated on top and the wiring simply printed. Hand-laid wires were no longer necessary — which is exactly what the next entry depends on.

People and institutions: Jean Hoerni / Fairchild Semiconductor

NotePatent US3025589 was filed May 1959 and granted 20 March 1962.

1959Jul 1959
Milestone

The form of the integrated circuit that could actually be manufactured

Robert Noyce patented a structure in which aluminium wiring is evaporated onto Hoerni's flat surface, with the oxide isolating the devices underneath. No hand assembly anywhere: the whole circuit is made by photographic printing and deposition. This fixed the method still in use today — convert a circuit diagram into a pattern and print it. Kilby had the idea; this patent made it something a factory could build.

People and institutions: Robert Noyce / Fairchild Semiconductor

NotePatent US2981877 was filed 30 July 1959 and granted 25 April 1961.

1960
Milestone

The MOS transistor — open the channel with a voltage alone

Kahng and Atalla placed a metal electrode over a silicon surface with an oxide layer between, and opened and closed the channel underneath using that electrode's voltage alone. Lilienfeld's 1926 patent had finally become real, 34 years later. Control by voltage rather than current means almost no power is drawn while idle, and the simple structure shrinks well. Nearly everything inside a chip today is this device.

  • After Lilienfeld's patent 34 years

People and institutions: Dawon Kahng / Mohamed Atalla / Bell Telephone Laboratories

NotePatent US3102230 was filed May 1960 and granted 27 August 1963 — 34 years after Lilienfeld's 1926 patent.

Connections to other timelines

Made possibleThe Automobile1997The Prius — the hybrid reaches mass production

In the first Prius, the heart of the inverter — which turns the battery's direct current into alternating current for the motor — was a power semiconductor called the IGBT. An IGBT switches large currents using only the voltage on an insulated gate, the same arrangement as the MOS transistor; Baliga and colleagues at GE described it in 1982. Control by voltage, the MOS idea, carried over into the heavy currents that move a car.

1963
Milestone

CMOS — a circuit that draws current only while it switches

Frank Wanlass and Chih-Tang Sah paired MOS transistors of opposite type. In either resting state one of the pair is closed, so no current runs from the supply to ground; power is consumed only during the instant of switching. Packing billions of devices onto one chip is possible only because each one draws essentially nothing while idle. No modern chip works without this arrangement.

  • Standby power nanowatts

People and institutions: Frank Wanlass / Chih-Tang Sah / Fairchild Semiconductor

NotePresented at ISSCC 1963; patent US3356858 was filed June 1963 and granted 5 December 1967.

1965Apr 1965
Theory

Moore's law — an observation that became the industry's target

In a magazine article Gordon Moore noted, from about five measured points, that the number of components on a chip had been doubling roughly every year, and said the trend would continue for some time. He revised it to every two years in 1975. The important thing is that this is not a law of physics. The industry planned its investment and its schedules around the line, and so it became self-fulfilling for half a century.

  • Data points behind the prediction five

People and institutions: Gordon Moore

NoteThe original prediction was a doubling every year, revised in 1975 to every two years. The often-quoted "every 18 months" is not Moore's own wording.

1968Jun 1968
Milestone

One bit, two parts

Robert Dennard devised a memory cell that stores a bit as charge on a capacitor, with a single transistor to put it in and take it out. The charge leaks away, so the cell has to be read and rewritten periodically — which is why it is called dynamic memory. Prevailing designs used six transistors per bit, so this cut the parts count by two-thirds, and the price of memory fell accordingly.

  • Parts per bit 1 transistor + 1 capacitor

People and institutions: Robert Dennard / IBM

NoteConceived in 1966; patent US3387286 was filed July 1967 and granted 4 June 1968.

1971Nov 1971
Deployment

The Intel 4004 — the computing part becomes a single component

Designed for a calculator made by the Japanese firm Busicom, the 4004 fitted the core of a processor onto one chip using 2,300 transistors, at a minimum feature size of 10 micrometres — about a tenth the width of a human hair. The key decision was to build a general-purpose processor that reads its instructions, rather than a circuit wired for calculators. Von Neumann's architecture was now available at the price of one component.

  • Transistors 2,300
  • Minimum feature size 10 µm

People and institutions: Federico Faggin / Marcian Hoff / Stanley Mazor / Masatoshi Shima (嶋正利) / Intel

Connections to other timelines

Made possibleThe Automobile1977A computer goes into the engine — the moment of ignition, decided by calculation

The first computer in a production car was the ignition controller on the 1977 Oldsmobile Toronado. Without the microprocessor putting the core of a computer onto a single component, there was no way to fit a computer into a cramped, hot engine bay. It was not the 4004 itself that went into the car, though.

1974–2003When Smaller Meant Faster and Cheaper
1974Oct 1974
Theory

Dennard scaling — why making things smaller made everything better

Dennard and colleagues showed that if a transistor's dimensions and its supply voltage are all reduced by the same factor, the device gets faster while power per unit area stays flat. Normally, packing parts closer together means a heat problem; not if the voltage comes down too. This was the justification for driving relentlessly toward smaller features. It is the free lunch referred to later on this timeline — and it ran out in the mid-2000s.

  • Shrink linear size to 0.7× half the area, 1.4× speed, same power density

People and institutions: Robert Dennard / IBM

Connections to other timelines

Made possibleNuclear Fusion2015Wendelstein 7-X first plasma — the stellarator line returns

Optimising a stellarator's magnetic field numerically, and designing the twisted coils that produce it, needed computing power that only became available in the 1980s. Dennard scaling — smaller devices that are faster, cheaper and less power-hungry — is what produced that computing power. W7-X is a machine that could only be designed once computers had caught up.

1980s1980s
Deployment

Japan takes the DRAM market — and it becomes a trade dispute

NEC, Hitachi, Toshiba and others outdid American makers on yield and quality, and by the mid-1980s Japanese firms are generally reckoned to have held around 80% of the world DRAM market. The 1986 US–Japan Semiconductor Agreement followed, putting prices and market access into the hands of governments. This is when semiconductors stopped being only a contest of performance and became a matter of state.

  • Japanese share of world DRAM (mid-1980s) ~80%

People and institutions: NEC (日本電気) / Hitachi (日立製作所) / Toshiba (東芝)

NoteJapanese makers' share of the world DRAM market is generally put near 80% in the mid-1980s, though figures vary by source. An NBER analysis chapter is cited here.

1987
Deployment

Separating the company that designs from the company that builds

TSMC, founded in Taiwan by Morris Chang, had no products of its own: it manufactured other companies' designs and nothing else. At the time owning a fab was simply what it meant to be a semiconductor company, so the idea looked eccentric. But as the cost of a fab climbed, design-only companies became viable. Nvidia and Qualcomm exist because this division of labour did.

People and institutions: Morris Chang (張忠謀) / TSMC

1997Dec 1997
Milestone

Switching the wiring to copper — the delay was in the connections, not the devices

Shrinking devices also thins the wires between them, and thin wires conduct badly. By the 1990s a chip's speed was set by the resistance of its wiring rather than by its transistors. IBM established a process using copper, which conducts better than aluminium. Copper had long been avoided because it contaminates silicon; enclosing each wire in a thin barrier layer solved that. Every fast chip since has copper wiring.

  • Wiring resistance vs aluminium ~40% lower

People and institutions: Daniel Edelstein / IBM

2003Dec 2003
Milestone

Stretching the silicon — the first trick that was not about size

At the 90 nm node Intel began deliberately stretching and compressing the silicon crystal. Shifting the spacing between atoms lets electrons move more easily, so current rises without any change in dimensions. This is the turning point. Until then, smaller had been enough on its own; from here on, each generation needed some additional contrivance beyond shrinking to get faster at all.

  • Gate length 45 nm (at the 90 nm node)

People and institutions: Tahir Ghani / Intel

2004–2022Scaling After the Free Lunch
2004
Discovery

Voltage stops coming down, and the free lunch ends

Dennard scaling assumed the supply voltage could fall along with the dimensions. But below a point, a transistor leaks current even when it is supposed to be off. If voltage cannot fall, packing more devices in means more heat. Clock frequencies stalled around 3–4 GHz, and makers stopped trying to make one core faster and started putting several side by side. The era in which smaller simply meant better was over.

  • Clock frequency plateaued around 3–4 GHz

NoteThe breakdown of Dennard scaling is usually placed around 2004–2006 and is not tied to a single event. 2004 is used here as the year the clock-frequency plateau became unmistakable; the 2011 paper that quantified the consequences is cited.

2007Dec 2007
Milestone

Abandoning the oxide after fifty years

The heart of a MOS transistor is the thin insulator between the gate electrode and the channel. Shrinking had thinned it to a few atoms, and electrons were tunnelling straight through. At the 45 nm node Intel replaced silicon dioxide with a hafnium-based material. Its higher permittivity means the same effect from a physically thicker film, and far less leakage. A material in use since the 1950s was dropped here for the first time.

  • Gate leakage reduced more than 10×

People and institutions: Kaizad Mistry / Intel

Read more on this site

The Rise of Quantum Mechanics, Part 2Out Through a Wall It Cannot Climb — Gamow and Quantum Tunnelling, 1928

Electrons leaking through a film only a few atoms thick is quantum tunnelling. This piece tells how the idea that a particle can turn up beyond a wall it cannot climb was born in 1928, from a puzzle about atomic nuclei.

2011May 2011
Milestone

The transistor stands up

While the channel stays a flat sheet, shrinking weakens the gate's grip on it until it can no longer be shut off cleanly. Intel raised the channel into a fin and wrapped the gate around three of its sides, putting the structure into volume production at the 22 nm node. More surrounding surface means a more decisive off. For fifty years chips had been built on a flat plane; here they started to grow upward.

  • Sides of the channel under control 1 → 3

People and institutions: Chris Auth / Intel / Digh Hisamoto (久本大)

NoteAnnounced May 2011, in volume production at the 22 nm node from 2012. The three-dimensional structure itself was anticipated by the DELTA device reported by Digh Hisamoto and colleagues at Hitachi (IEDM 1989).

2019Oct 2019
Deployment

Extreme ultraviolet lithography finally reaches the factory

Printing finer patterns needs shorter wavelengths, but light at 13.5 nm is absorbed by every material, glass lenses included. So the machine is built entirely from mirrors, its interior held in vacuum, and its light made by firing a laser at droplets of tin. TSMC shipped volume products from an EUV process in 2019. More than twenty years passed between the start of the research and that shipment — one of the longest gaps on this timeline.

  • Wavelength 193 nm → 13.5 nm

People and institutions: TSMC / ASML / Samsung

NoteHow long EUV took depends on where you start counting; from the industry-academia consortia of the mid-1990s it is more than twenty years.

Connections to other timelines

Shares a basis withNuclear Fusion2022Ignition — the Lawson criterion is crossed for the first time

The EUV light source and laser fusion rest on the same physics: fire an intense laser at a tiny target and make a hot plasma for an instant. The EUV source is designed with radiation-hydrodynamics simulations that follow how that plasma expands and how much light it gives off. In Japan, researchers at Osaka University's Institute of Laser Engineering — the country's centre for laser fusion — used this method to pin down the conditions under which tin plasmas emit efficiently, and published the work in a fusion research journal.

2022Jun 2022
Deployment

Wrapping the channel completely — gate-all-around goes into production

At its 3 nm node Samsung began volume production of a structure in which the channel is a stack of thin sheets with the gate wrapped entirely around each one. Four sides rather than three means a still more decisive off, and varying the number and width of the sheets lets a designer tune a device without changing the process. The move away from fins is also a sign that shrinking has nearly exhausted the tricks available beyond size.

  • Sides of the channel under control 3 → all around

People and institutions: Samsung / Nicolas Loubet / IBM

2025–2035Down to Counting Atoms
2025
Deployment

TSMC enters volume production at the 2 nm node

TSMC began volume production of N2 in the fourth quarter of 2025, its first generation using gate-all-around transistors. It is worth being clear that "2 nm" is no longer a dimension: nothing in the device measures 2 nm, and the number is a name used to distinguish generations. What is actually at stake is how many devices fit into a given area.

  • Volume production Q4 2025
  • Transistor first-generation nanosheet (gate-all-around)

People and institutions: TSMC

NoteTSMC states that N2 entered volume production in 4Q25. "2 nm" is a node name, not a dimension that exists anywhere in the device.

2025Oct 2025
Deployment

Delivering power from the back of the wafer

A chip's front side carries both signal wiring and power wiring, and the two get in each other's way. At its 18A node Intel thinned the wafer and fed power in from the back, leaving the front for signals alone. The freed-up wiring space allows denser packing, and the voltage droop under load is reduced. Introducing gate-all-around and backside power in the same generation makes this a large structural change by any standard.

  • Where power comes from front side → back side

People and institutions: Intel

NoteThe reported benefits of backside power (density, voltage droop) come from Intel's own VLSI 2023 test-chip results. Independent confirmation of 18A volume figures is still pending.

2026
Deployment

The market approaches a trillion dollars — pulled by AI

World Semiconductor Trade Statistics puts the 2025 market at roughly USD 772 billion and forecasts about USD 975 billion for 2026, with logic chips for computation and the memory attached to them driving most of the growth. It is 180 years since Boole turned logic into algebra. A forecast is only a forecast, though: this is the entry on the timeline that will go stale first.

  • World market, 2025 ~USD 772 bn
  • 2026 forecast ~USD 975 bn

People and institutions: WSTS

NoteThe 2026 figure of USD 975 billion is the WSTS autumn 2025 forecast, not an actual result. Replace it when the next forecast is issued.

Connections to other timelines

Under the same pressure asNuclear Fusion2025Electricity from a plant that does not yet exist is bought and sold

Both sit behind the same demand for AI computation. Chips for computation are what is driving the semiconductor market's growth, and in announcing their deal, CFS and Google named rising electricity demand from AI and other advanced computing among the needs fusion power would help meet. The industry that makes the computers and the search for power to run them are under the same pressure at the same moment.

2027c.
Milestone

High-NA EUV — a machine costing over 300 million dollars

Numerical aperture measures how wide an angle of light the optics can gather; raising it from 0.33 to 0.55 allows finer lines to be printed. The machine gets larger, and a single unit is reported to cost over USD 300 million. Intel plans to use it at its 14A node, though the timing of volume use is not settled. Tools like this are why the cost of a fab is now discussed alongside national budgets.

  • Numerical aperture 0.33 → 0.55
  • Price per machine (reported) over USD 300 m

People and institutions: ASML / Intel

NoteIntel has indicated it will use High-NA EUV (the ASML EXE:5200 family) at its 14A node, but the timing of volume use is not fixed. Keep this entry marked projected until it resolves.

2033c.
Milestone

What comes next — stacking the two transistor types on top of each other

CMOS uses transistors of opposite type in pairs, and today they sit side by side on a plane. Stack one above the other and the area per pair roughly halves. The research institute imec places commercial introduction around 2033. Building upward because there is no more room sideways continues the logic of fins and gate-all-around, and shows plainly that shrinking has become a history of giving up on the flat plane.

  • n- and p-type devices side by side → stacked vertically

People and institutions: imec

Noteimec's roadmap as of 2026 places commercial introduction of CFET around 2033, at the A7 node. The roadmap is revised annually; replace this when a new version appears.