Light, semiconductors and information: from the transistor to the network: The transistor and the integrated circuit
All courses / Understanding Engineering / Module 9: Light, semiconductors and information: from the transistor to the network Part 4 of 5The transistor and the integrated circuit
One tiny device, the transistor, is a switch and an amplifier with no moving part, and printing it by the billion on one crystal made computing cheap. This part follows it from gate to logic, from sand to chip, through Moore's law and its cost estimate, and ends at the energy and heat that now limit how fast chips can go.
4.1 A transistor lets a voltage on one terminal control a current between two others, so it can amplify or switch
The transistor in nearly every chip is the MOSFET. A thin channel of silicon joins two terminals, the source and the drain, and above the channel sits a third, the gate, separated from it by an insulating oxide only a few atoms thick. With the gate at zero the channel has no free carriers and is shut; a gate voltage draws carriers in by its electric field and opens it. No part moves, and the insulated gate takes almost no steady current, so one transistor can drive many others.
Hold the gate part-way open and the transistor becomes an amplifier: a small change in gate voltage makes a larger change in drain current. A shop's electronic weighing scale needs this: its load cell, a metal beam with strain gauges bonded to it, gives only a few millivolts at full load. Say the first stage passes 10 mA more drain current per volt on its gate. A 10 mV change from the load cell then shifts the current by 0.1 mA, and through a 2,000 ohm resistor that is a 0.2 V swing, a voltage gain of 20. The added power comes from the scale's battery; the load cell only steers it.
Drive the gate fully on or fully off instead and the transistor is a switch, with an on-to-off current ratio that can pass a million. CMOS logic pairs an n-type transistor, which pulls the output down, with a p-type one, which pulls it up. A high input gives a low output and a low input a high one: an inverter. In a steady state one of the pair is always off, so current flows mainly at the moment of switching. Four transistors make a NAND gate, and NAND gates alone can build any logic.
Before the transistor (1947) these jobs fell to vacuum tubes, which boil electrons off a hot filament; ENIAC, completed in 1945, used about 18,000 of them and drew roughly 150 kW.
- Bloomfield, How Things Work (Wiley, 2015) · ch. 14, 'Optics and Electronics', section 'Audio Players' (transistors, MOSFETs, logic elements, amplifiers)
4.2 Printing with light makes a billion transistors in the same few steps that make one
The planar process is the idea behind the integrated circuit. Start with a wafer, a thin slice of a single silicon crystal, 300 mm across. Grow an oxide skin, coat it with photoresist, a chemical that changes in light, and shine light through a mask that carries the pattern. Wash away the exposed resist, etch or implant dopants through the openings, strip the resist and repeat. A modern logic chip needs more than 60 such rounds, transistors first and copper wiring above.
Each round treats the whole wafer at once, so a pattern with a billion features costs about the same as one with a thousand. A chip's price depends on its area and the number of steps, not on its transistor count. Halve each dimension and the same wafer holds four times as many transistors for almost the same price.
How small can the pattern be? Light spreads by diffraction, as Part 2 showed for a lens, so the smallest printable half-pitch is about 0.3 x wavelength / numerical aperture. Deep ultraviolet light of 193 nm through immersion optics (NA 1.35) gives 0.3 x 193 / 1.35 = 43 nm in one exposure. Extreme ultraviolet of 13.5 nm with NA 0.33 gives 0.3 x 13.5 / 0.33 = 12 nm. That explains the shift to EUV machines, which use mirrors in vacuum because glass absorbs this light. Chip names such as '3 nm' are now labels, not real distances.
carbon strips the oxygen: metallurgical silicon, about 98 to 99 per cent pure
2 Purification to polysiliconmade into a gas, distilled, and redeposited: purity of 99.9999999 per cent or better
3 Crystal growth and slicingone crystal pulled from a melt, sawn and polished into 300 mm wafers
4 Fab: 60 or more print roundsoxide, resist, light, etch, dopant, metal, repeated; each round treats the whole wafer
5 Test, saw and packagegood dies cut out, wired and sealed
Outputs Chipshundreds per wafer, each with billions of transistorsSite pages: 'Silicon', 'Polysilicon', 'Silicon wafer', 'Planar process and photolithography', 'Photoresists', 'Extreme ultraviolet lithography', 'Semiconductor manufacturing equipment', 'Integrated circuit (chip)'.
Source: Bloomfield, How Things Work (Wiley, 2015), ch. 14, 'Optics and Electronics' (diffraction limit, transistors); process description from the author's own knowledge- Bloomfield, How Things Work (Wiley, 2015) · ch. 14, 'Optics and Electronics' (diffraction limit)
4.3 Moore's law is an observed doubling every two years, and the estimate of a transistor's price shows why it mattered
In 1965 Gordon Moore noticed that the components on the cheapest chip were doubling about yearly, and later revised this to every two years. It is a pattern held by an industry that set itself a schedule, not a law of nature. The first microprocessor, Intel's 4004 of 1971, held 2,300 transistors on about 12 mm2. From 1971 to 2021 is 50 years, or 25 doublings, and 2 to the power 25 is about 34 million. 2,300 x 34 million is 8 x 10^10. Nvidia's H100 processor of 2022, one of the largest single chips made, holds about 80 billion.
Now the price. Take a 300 mm wafer: its area is pi x 150 mm x 150 mm, about 70,700 mm2. A 100 mm2 chip gives roughly 700 sites, 600 after losing the ragged edge, and with 80 per cent of them working, 480 good chips. Assume a processed wafer costs ₹15 lakh, about US$17,000; this is an assumed round figure for an advanced node, not a quoted price. One good chip then costs ₹3,100. At a density of 100 million transistors per mm2, a round figure for leading logic, it holds 10 billion, so a rupee buys about three million transistors. Density grew far faster than wafer prices, and that is the whole story.
The real effect was cheapness: functions once kept for a mainframe now sit in a phone or a ceiling fan's controller. The price of the fab has risen with each step, though, and only a handful of firms can afford the leading ones.
| Step | Working | Result |
|---|---|---|
| Doublings, 1971 to 2021 | 50 years / 2 | 25 |
| Growth factor | 2^25 | about 34 million |
| Predicted transistors | 2,300 x 34 million | about 8 x 10^10 |
| A large real chip, 2022 (Nvidia H100) | maker's figure | about 80 billion |
| Wafer area | 3.14 x 150 x 150 mm2 | 70,700 mm2 |
| Chips of 100 mm2 per wafer | 707 minus edge loss | about 600 |
| Good chips at 80 per cent yield | 600 x 0.8 | 480 |
| Cost per good chip (wafer ₹15 lakh, assumed) | ₹15,00,000 / 480 | about ₹3,100 |
| Transistors per chip at 100 million per mm2 | 100 mm2 x 100 million | 10 billion |
| Transistors per rupee | 10 billion / ₹3,100 | about 3 million |
All values calculated. The wafer price and density are round assumptions for the scale of the answer, not quotations. The doubling period is an observed average and has stretched in recent years.
Source: Bloomfield, How Things Work (Wiley, 2015), ch. 14, 'Optics and Electronics' (transistors, logic elements); arithmetic by the author- Bloomfield, How Things Work (Wiley, 2015) · ch. 14, 'Optics and Electronics' (integrated circuits, bits and bytes)
- Moore, G. E., 'Cramming More Components onto Integrated Circuits', Electronics 38(8), 1965 · the observed doubling of components per chip
- NVIDIA, 'NVIDIA H100 Tensor Core GPU Architecture' white paper (2022) · about 80 billion transistors on one H100 chip
4.4 Every switch costs a little energy, and a billion of them add up to a hotplate's heat in a fingernail of silicon
A CMOS gate draws energy when it switches, because it must charge and discharge its own capacitance, and the energy per switch is of the order of half x capacitance x voltage squared. For an advanced transistor with its wiring, take 0.1 fF (10^-16 farad) and 0.7 V. That gives 0.5 x 10^-16 x 0.49, about 2.5 x 10^-17 joule. Chip power = transistors x share switching each cycle x energy per switch x clock rate. With 10^10 transistors, 5 per cent active, and 3 GHz, that is 1.5 x 10^18 switches a second x 2.5 x 10^-17 J, about 38 W, which is the real range of a laptop or phone processor.
The area is the trouble. Those 38 W leave a 1 cm2 die: 38 W per cm2. A 1,500 W induction cooker spreads its heat over some 250 cm2, 6 W per cm2. The chip is six times hotter per area, which is why heat sinks, fans and a phone's thermal throttling are part of the design.
For decades shrinking avoided this: when dimensions and voltage fell together, power per area stayed fixed (Dennard scaling). It ended around 2005, because voltage could not fall further. A transistor can turn off no faster than a factor of ten per 60 mV of gate voltage, so a million-to-one on-off ratio needs about 0.36 V, plus margin. Clock speeds stopped near 3 to 5 GHz, and chips turned to more cores and specialised units.
Is there a floor? Erasing one bit of information must release at least kT x ln 2, the Landauer limit, 2.9 x 10^-21 J at room temperature. Today's switches use about 9,000 times that, so physics leaves room, but wires, leakage and cost leave less. And silicon's lattice repeats every 0.54 nm, so a few-nanometre feature is only some atoms across.
| Step | Working | Result |
|---|---|---|
| Energy per switch | 0.5 x 0.1 fF x (0.7 V)^2 | about 2.5 x 10^-17 J |
| Switches per second | 10^10 transistors x 5% x 3 GHz | 1.5 x 10^18 |
| Chip power | 1.5 x 10^18 x 2.5 x 10^-17 J | about 38 W |
| Power per area, 1 cm2 die | 38 W / 1 cm2 | about 38 W/cm2 |
| Induction cooker, 1,500 W on 250 cm2 | 1,500 / 250 | about 6 W/cm2 |
| Landauer limit at 300 K | 1.38 x 10^-23 x 300 x 0.69 | 2.9 x 10^-21 J |
| Switch energy against the limit | 2.5 x 10^-17 / 2.9 x 10^-21 | about 9,000 times |
All values calculated. Capacitance, activity and clock rate are round figures to show the scale; real chips differ by design. The 60 mV per tenfold current is a room-temperature limit of this kind of switch.
Source: Landauer, R., 'Irreversibility and Heat Generation in the Computing Process', IBM Journal of Research and Development 5(3), 1961, minimum energy to erase a bit, kT ln 2; arithmetic by the author- Bloomfield, How Things Work (Wiley, 2015) · ch. 14, 'Optics and Electronics' (transistors, MOSFETs, bits and bytes)
See it in the table
Cu CopperSi SiliconCheck your understanding
Why does an integrated circuit with a billion transistors cost about as much as one with a million of the same area?
Light prints the pattern in parallel. Shrinking features puts more transistors in the same area for nearly the same processing cost.A chip has 10^10 transistors, 5 per cent switching each cycle at 3 GHz, 2.5 x 10^-17 J per switch. About what is its power?
Switches per second are 10^10 x 0.05 x 3 x 10^9 = 1.5 x 10^18; times 2.5 x 10^-17 J gives about 38 W.Why did clock speeds stop climbing near 3 to 5 GHz after about 2005?
The end of Dennard scaling: a transistor needs a minimum gate swing to turn off, so voltage stalled, and faster clocks meant more heat per area than cooling could remove.