Sixty years of shrinking
The first integrated circuits of the early 1960s held a handful of transistors. A large processor or AI accelerator of 2026 holds more than a hundred billion, each switching billions of times a second, in a few square centimetres of silicon. No other manufactured thing has improved by a factor of ten billion in a working lifetime.
That improvement was not a law of nature but a plan, repeatedly declared dead and repeatedly rescued by new physics, materials, machines and money. You met the MOSFET as a switch in Transistors. This lesson asks what happens to it when you make it smaller, why the easy part ended around 2005, and how the industry kept going: new transistor shapes (FinFETs, then gate-all-around nanosheets, in production at the 2 nm class in 2025–26), power wires moved to the back of the wafer, and light of 13.5 nm wavelength made by vaporising droplets of tin. It ends with money: a leading-edge fab costs tens of billions of US dollars, a fact that shapes the design choices of How chips are designed and Hardware for AI.
Moore's law and Dennard scaling
Definition— Moore's law
In 1965 Gordon Moore, then at Fairchild Semiconductor, observed that the number of components on the most cost-effective integrated circuit had doubled every year, and predicted a decade more of it (Moore, 1965); in 1975 he revised the rate to every two years. Moore's law now usually means: the number of transistors on a chip of a given cost doubles roughly every two years — an economic observation, not a law of physics.
Moore's law says how many, not how fast or how hot. That was answered in 1974 by Robert Dennard and colleagues at IBM (Dennard et al., 1974): shrink everything together — lengths, oxide thickness and voltage — so the electric fields inside the device stay the same.
Definition— Constant-field (Dennard) scaling
Scale every linear dimension of a MOSFET, and its supply and threshold voltages, by (with ), and raise the doping by . Then, to first order:
| Quantity | Scales as |
|---|---|
| Area per transistor | |
| Gate capacitance | |
| Voltage , current | |
| Gate delay | , so clock frequency can rise by |
| Dynamic power per transistor | |
| Power density (power per area) | 1: unchanged |
The last line is the miracle. A "generation" was : twice the transistors per area, each about 40% faster, and the chip's power per square centimetre the same as before. For thirty years every new process gave more, faster, and cooler-per-transistor circuits at once.
Theorem— Power density under constant-field scaling
If dimensions and voltages scale by and frequency by , with the switching activity fixed, the dynamic power per unit area is unchanged.
Proof
Dynamic power per transistor is . Under scaling it becomes . The number of transistors per unit area grows by , so power per area is the original value.
∎
Why it ended
The proof needs to fall by every generation. Supply voltages fell from 5 V in the 1980s to around 1 V by the mid-2000s, and then mostly stopped. To keep the transistor fast, the overdrive must stay healthy, so the threshold voltage must fall with . But below threshold a MOSFET does not switch off cleanly: its current falls exponentially with gate voltage, one decade for every millivolts, where the subthreshold swing is
The bound comes from the Boltzmann distribution of electron energies, not from geometry. Every 60–80 mV taken off multiplies every transistor's off-state leakage by ten; with billions of transistors, leakage power rivalled switching power, and — and with it — could fall no further.
import math
k = math.sqrt(2) # one "generation": lengths shrink by 1/k, so area per transistor halves
def generations(rule, n=6):
"""Normalised capacitance C, supply V, clock f, transistors per area, and power density."""
C, V, f, density = 1.0, 1.0, 1.0, 1.0
rows = []
for g in range(n + 1):
power_per_transistor = C * V**2 * f # dynamic power, alpha * C * V^2 * f, alpha fixed
rows.append((g, V, f, density, density * power_per_transistor))
C /= k
density *= k**2
if rule == "Dennard":
V /= k
f *= k
elif rule == "V fixed, f rising":
f *= k
# "V and f fixed": nothing else changes (roughly the post-2005 regime)
return rows
for rule in ["Dennard", "V fixed, f rising", "V and f fixed"]:
print(f"{rule}:")
print(" gen V f transistors/area power density")
for g, V, f, d, pd in generations(rule):
print(f" {g:3d} {V:5.2f} {f:5.2f} {d:10.0f} {pd:8.2f}")
# Why V stopped scaling: the subthreshold swing.
kT_q = 1.380649e-23 * 300 / 1.602176634e-19 # thermal voltage at 300 K, in volts
S_min = math.log(10) * kT_q
print(f"\nthermal voltage kT/q at 300 K = {kT_q*1000:.1f} mV")
print(f"ideal subthreshold swing ln(10) kT/q = {S_min*1000:.1f} mV per decade of current")
for S_mV in (60, 70, 80):
for dVt_mV in (100, 200):
print(f" S = {S_mV} mV/dec: lowering Vt by {dVt_mV} mV multiplies leakage by {10**(dVt_mV / S_mV):8.0f}")Under Dennard's rules power density stays at 1.00 for six generations. Freeze the voltage but keep raising the clock, and it doubles every generation, 64 times after six. Desktop processors of the mid-2000s already dissipated of order 100 W per square centimetre, more than ten times a kitchen hotplate, so that path was closed. The industry took the third route: hold the clock near 3–5 GHz, accept power density rising by about per generation, and spend the extra transistors on more cores, bigger caches and specialised accelerators. Even so, not every transistor can switch at full speed at once — the "dark silicon" problem (Esmaeilzadeh et al., 2011), and one reason AI accelerators exist.
From planar to FinFET to gate-all-around
If cannot fall, the other lever is to make the transistor turn off more sharply — bring close to its 60 mV ideal — while still shrinking it. In a short planar transistor the drain's field reaches under the gate and lowers the barrier (drain-induced barrier lowering, DIBL), so the drain partly controls the current. The cure is geometric: surround the channel with gate so the gate's field dominates everywhere in the silicon.
Planar MOSFETs carried the industry to about the 28–20 nm generations, helped by strained silicon (higher mobility, from around 2003) and high-k dielectrics with metal gates (high gate capacitance without tunnelling through an oxide a few atoms thick, from 2007).
FinFETs stand the channel up as a thin vertical fin and wrap the gate over its top and both sides. The idea was developed in the late 1990s at Hitachi and the University of California, Berkeley (Hisamoto et al., 2000). Intel put it in production at 22 nm in 2011–12 (as its "tri-gate"), TSMC and Samsung at their 16 and 14 nm processes around 2014–15. A fin only a few nanometres thick has no silicon far from the gate, so DIBL and leakage fall sharply. The price is quantised width: a transistor's drive current is set by how many fins it has — one, two, three — so designers cannot choose an in-between size.
Gate-all-around (GAA) nanosheets turn the fin on its side and slice it: a stack of typically three or four thin horizontal sheets, each completely surrounded by gate (Loubet et al., 2017). Full wrap-around gives the best electrostatic control a silicon device can have, and the sheet width can be drawn at any value, so the designer gets back a continuous knob between drive strength and area. As of late 2026:
| Company | Gate-all-around process | Status |
|---|---|---|
| Samsung | SF3 ("3 nm GAA", MBCFET), then SF2 | First GAA in production, from 2022; SF2 in volume production from late 2025 (the Exynos 2600) |
| TSMC | N2, then N2P and A16 | N2 in volume production since the fourth quarter of 2025; N2P and A16 scheduled for the second half of 2026 |
| Intel | 18A ("RibbonFET", with PowerVia) | High-volume manufacturing in Arizona from 2025; first products (Core Ultra series 3, "Panther Lake") from early 2026 |
| Rapidus (Japan) | 2 nm, developed with IBM from its 2021 research process | Pilot line in Chitose, Hokkaido, made its first GAA transistors in 2025; mass production targeted for 2027 |
These statuses come from the companies' own announcements (TSMC, Intel, Samsung, Rapidus) and change quickly; yields and performance claims are not independently published.
Making a nanosheet is far harder than drawing one: the stack starts as alternating epitaxial layers of silicon and silicon–germanium, the SiGe is etched away sideways to leave silicon sheets suspended like the rungs of a ladder, and the gate dielectric and metal are then deposited, atomic layer by atomic layer, into gaps a few nanometres high.
Backside power delivery
A chip has two wiring jobs. Signals need many thin, dense wires; power needs thick, low-resistance wires carrying large currents with little voltage drop. Traditionally both share one stack of a dozen or more copper layers above the transistors, and power descends from bumps on top through every layer, each thinner and more resistive than the last. That costs voltage (the IR drop, which you will compute at rack scale in Powering the AI data centre) and routing tracks that signals could have used.
Definition— Backside power delivery network (BSPDN)
A wafer-processing scheme in which the power wiring is built on the back of the wafer: after the transistors and the front-side signal wiring are made, the wafer is bonded face-down to a carrier, the original silicon is thinned to within a few hundred nanometres of the transistors, and nanoscale vias and a thick power network are built from the back. The front side is then left for signals.
Intel was first to put backside power into high-volume production, as PowerVia in 18A; Intel's own figures claim much lower worst-case voltage droop and a block-level area saving of around 10% (Intel 18A). TSMC's Super Power Rail, which contacts sources and drains directly from the back, is part of its A16 process, scheduled for volume production in the second half of 2026; TSMC will keep offering front-side-power processes too, since backside processing adds cost. The new problems are real: thinning a bonded wafer to within a few hundred nanometres of the transistors without distorting the pattern, getting heat out through bonded layers, and failure analysis, which for twenty years has probed failing chips through the back.
What “2 nm” means
Nothing on a "2 nm" chip measures 2 nm. Until the late 1990s a node's name roughly matched the gate length or the half-pitch of the tightest metal lines. Then the names kept shrinking by the traditional 0.7 per generation while the physical dimensions shrank more slowly, and since FinFETs the name has been a label for a generation, chosen by each company. Intel renamed its processes in 2021 ("Intel 7", "Intel 4", "Intel 3", "20A", "18A") partly to align with competitors of similar density; TSMC's "A16" and "A14" count in ångströms.
What determines density is a handful of pitches — the repeat distance of a pattern, line plus space:
- Contacted gate pitch (CPP): the distance from one transistor gate to the next, including room for a contact between them. In the 3 to 2 nm classes it is in the mid-40s of nanometres.
- Minimum metal pitch (MP): the tightest wiring layer, in the low 20s of nanometres.
- Cell height: standard logic cells (the gates from Logic gates and Boolean algebra, laid out as fixed-height tiles) are a whole number of metal tracks tall, five or six in the densest libraries.
The IRDS roadmap names nodes by such numbers: G48M24 means a contacted gate pitch of 48 nm and a metal pitch of 24 nm (IRDS). Exact design rules are confidential, and reported densities depend on which cells are counted; for the densest 2 nm-class logic libraries, published estimates are a few hundred million transistors per square millimetre. You will estimate this from pitches in the next section.
Misconception
"A 2 nm process beats a 3 nm process by the ratio 3/2." The names are not measurements, and different companies' "2 nm" processes are not equal. Compare published figures for density, speed at a given power, and power at a given speed, measured on the same kind of circuit — and remember that SRAM, analog circuits and I/O shrink far less than logic.
Printing the pitches: EUV and high-NA EUV
Every pattern on a chip is printed by optical lithography: light passes through (or reflects from) a mask carrying the pattern, a lens system demagnifies it four times, and the image exposes a light-sensitive resist on the wafer. The smallest pitch a lens can print is set by diffraction. The Rayleigh criterion gives the smallest printable half-pitch as
where is the wavelength, NA the numerical aperture of the lens (the sine of its acceptance half-angle, times the refractive index of the medium), and a process factor that cannot go below 0.25 for a single exposure and is typically 0.3 to 0.4 in production.
From the mid-2000s the workhorse was ArF immersion lithography: 193 nm light, with water between lens and wafer to push NA to 1.35. Below its limit, fabs split each dense layer into two, three or four masks (multi-patterning), at great cost in steps, alignment error and time.
Extreme ultraviolet (EUV) lithography uses 13.5 nm light. Even air absorbs it, so the optical path is in vacuum and every element, mask included, is a mirror coated with dozens of molybdenum–silicon layers that reflect about two-thirds of the light per bounce. The source is a plasma: tens of thousands of droplets of molten tin per second, each struck by a pulse from a high-power CO₂ laser. The only supplier of EUV scanners is ASML (the Netherlands), with optics from Zeiss and drive lasers from Trumpf (both Germany) (ASML). EUV entered high-volume logic production around 2019 and now prints the critical layers of every leading-edge logic process and, increasingly, of DRAM.
High-NA EUV raises the numerical aperture from 0.33 to 0.55. Its anamorphic optics demagnify four times in one direction and eight in the other, which halves the exposure field from 26 × 33 mm to 26 × 16.5 mm. Intel installed the first production high-NA scanners and designed its 14A process around them, with a standard-EUV fallback; TSMC has said publicly that it does not need high-NA for A14. Reported prices are several hundred million euros per tool, about twice a standard EUV scanner, so the argument is cost per layer: one high-NA exposure against two cheaper ones.
import math
# 1. Resolution: the smallest printable half-pitch is about k1 * wavelength / NA.
tools = [("ArF immersion (DUV)", 193.0, 1.35),
("EUV, NA 0.33", 13.5, 0.33),
("High-NA EUV, NA 0.55", 13.5, 0.55)]
print("tool wavelength NA min pitch (k1=0.25) practical pitch (k1=0.35)")
for name, lam, na in tools:
p_limit = 2 * 0.25 * lam / na
p_practical = 2 * 0.35 * lam / na
print(f"{name:22s} {lam:6.1f} nm {na:4.2f} {p_limit:8.1f} nm {p_practical:8.1f} nm")
# 2. Photon shot noise: how many photons land on a small square of resist?
h, c, e = 6.62607015e-34, 2.99792458e8, 1.602176634e-19
dose = 30e-3 / 1e-4 # 30 mJ/cm^2, in J/m^2
for name, lam_nm in [("ArF 193 nm", 193.0), ("EUV 13.5 nm", 13.5)]:
E_photon = h * c / (lam_nm * 1e-9)
per_nm2 = dose / E_photon * 1e-18
n = per_nm2 * 10 * 10 # photons in a 10 nm x 10 nm square
print(f"{name}: photon energy {E_photon / e:5.1f} eV, {per_nm2:6.1f} photons/nm^2, "
f"{n:7.0f} in a 10x10 nm square, shot noise {100 / math.sqrt(n):4.1f}%")
# 3. Density from pitches, not from the node name.
def nand2_density(cpp_nm, mp_nm, tracks, cpp_per_cell=3):
"""Million transistors per mm^2 if a chip were tiled with 2-input NAND cells (4 transistors)."""
width = cpp_per_cell * cpp_nm # cell width: gate pitches across
height = tracks * mp_nm # cell height: metal tracks tall
area_mm2 = width * height * 1e-12 # nm^2 -> mm^2
return 4 / area_mm2 / 1e6
print("\nillustrative pitches for each class -> NAND2-cell density")
for label, cpp, mp, tracks in [("16/14 nm-class FinFET", 90, 64, 9),
("7 nm-class FinFET", 57, 40, 6),
("3 nm-class FinFET", 48, 24, 6),
("2 nm-class nanosheet", 45, 22, 5)]:
print(f" {label:22s} CPP {cpp} nm, MP {mp} nm, {tracks} tracks: {nand2_density(cpp, mp, tracks):6.0f} MTr/mm^2")
print(f" if '2 nm' were really both pitches: {nand2_density(2, 2, 5):,.0f} MTr/mm^2")Three things come out. First, single-exposure 0.33-NA EUV prints pitches down to roughly 28 nm in practice, so the tightest metal layers at 3 and 2 nm (low 20s of nm) need two EUV exposures or high-NA. Second, EUV photons carry 14 times the energy of ArF photons, so at the same dose there are 14 times fewer: about 2,000 in a 10 × 10 nm square, with a random (Poisson) fluctuation of a couple of per cent. These stochastic effects cause rough edges, missing contacts and bridged lines; the cures are a higher dose, which slows the scanner, and better resists. Third, the density estimate from pitches lands at a few hundred million transistors per mm² for the 2 nm class, the right order of magnitude, while a literal "2 nm" would give an absurd 67 billion.
Memory scales differently: DRAM and NAND
DRAM stores a bit as charge on a capacitor, one transistor and one capacitor per cell. The capacitor must still be readable after tens of milliseconds of leakage, so its capacitance cannot shrink with the transistor: it has become a tall, narrow cylinder many times higher than it is wide. DRAM has advanced by a nanometre or so per generation through the "10 nm class" (generations called 1α, 1β, 1γ, or 1a, 1b, 1c), and by 2025–26 Samsung and SK hynix (Korea) and Micron (United States) all use EUV for critical layers, alongside fast-growing Chinese makers. Next come vertical-channel (4F²) cells and, further out, 3D DRAM with cells stacked like NAND; stacked DRAM dies are already routine in the high-bandwidth memory of Hardware for AI.
NAND flash stopped shrinking in the plane around 15 nm, when so few electrons remained per cell that losing a handful corrupted the data. The industry turned the problem through 90 degrees: 3D NAND builds vertical strings of cells through a stack of word-line layers, so density grows by adding layers — mostly cheap deposition steps — rather than by lithography at the edge of the possible. Products with more than 300 layers were in production by 2025 (makers count layers differently), built as two or three stacked decks because etching one hole through the whole stack, at aspect ratios beyond 50:1, is too hard. Storing three or four bits per cell (TLC, QLC) adds density at a cost in endurance and speed.
The economics of fabs
Scaling continues only while it pays. Three numbers organise the economics.
The fab. A leading-edge logic fab — clean rooms the size of several football pitches, dozens of EUV scanners, thousands of other tools — costs of order USD 20 billion or more; TSMC alone has announced a planned USD 165 billion for its Arizona site. The tools are depreciated over about five years, so a fab must run full from the day it opens.
The wafer. A leading-edge 300 mm wafer passes through well over a thousand process steps in two to three months. Prices are not published; press reports suggest of order USD 20,000–30,000 for 3 and 2 nm-class processes.
Yield. Random defects kill dies. If killer defects land independently at an average density per cm², the number on a die of area is Poisson-distributed, and the fraction of dies with none is
Yield falls exponentially with die area.
import math
WAFER_DIAMETER = 300.0 # mm
def dies_per_wafer(die_area_mm2, d=WAFER_DIAMETER):
"""A standard approximation: wafer area / die area, minus the partial dies lost at the edge."""
return int(math.pi * (d / 2) ** 2 / die_area_mm2 - math.pi * d / math.sqrt(2 * die_area_mm2))
def poisson_yield(die_area_mm2, d0_per_cm2):
"""Fraction of dies with no killer defect, if defects land independently at random."""
return math.exp(-die_area_mm2 / 100 * d0_per_cm2)
def cost_per_good_die(wafer_cost, die_area_mm2, d0_per_cm2):
good = dies_per_wafer(die_area_mm2) * poisson_yield(die_area_mm2, d0_per_cm2)
return wafer_cost / good
WAFER_COST = 25_000 # USD; an assumed round number for a leading-edge wafer (real prices are not published)
print("die area dies/wafer yield (D0=0.1/cm^2) cost per good die cost per mm^2 of good silicon")
for area in (25, 100, 300, 600, 800):
n = dies_per_wafer(area)
y = poisson_yield(area, 0.1)
cpd = cost_per_good_die(WAFER_COST, area, 0.1)
print(f"{area:5d} mm^2 {n:6d} {y:6.1%} USD {cpd:8,.0f} USD {cpd / area:5.2f}")
print("\nsame 600 mm^2 die while the process matures (defect density falling):")
for d0 in (0.5, 0.3, 0.2, 0.1, 0.05):
print(f" D0 = {d0:4.2f}/cm^2: yield {poisson_yield(600, d0):6.1%}, cost per good die USD {cost_per_good_die(WAFER_COST, 600, d0):9,.0f}")
# Spreading the fab: a USD 20 billion fab written off over 5 years, running 50,000 wafers a month.
fab, years, wafers_per_month = 20e9, 5, 50_000
print(f"\ndepreciation alone: USD {fab / (years * 12 * wafers_per_month):,.0f} per wafer")Read the first table down the last column: a square millimetre of good silicon costs almost three times as much in an 800 mm² die as in a 25 mm² one, because big dies both waste the wafer's edge and catch more defects. The second table shows why a new process is so expensive in its first year: the same 600 mm² die costs over ten times as much at as at . And depreciation alone, for an assumed USD 20 billion fab running 50,000 wafers a month, is several thousand dollars per wafer before any materials, labour or energy.
These numbers explain three features of the 2026 industry. Large AI chips are built from several smaller dies (chiplets) in one package. Only a few companies — TSMC (Taiwan, with newer fabs in Arizona and Japan and one under construction in Germany), Samsung (Korea, with a fab in Texas) and Intel (United States, Ireland and Israel), with Rapidus aiming to join and China's SMIC producing 7 nm-class chips with DUV multi-patterning — operate at or near the leading edge. And governments subsidise fabs heavily (the United States' CHIPS and Science Act, the European Chips Act, and programmes in Japan, Korea, China and India) because the supply chain is concentrated in so few places.
After 2026: forksheets, CFETs and 2D channels
Papers at the IEEE International Electron Devices Meeting (IEDM) and the VLSI Symposium, from imec, IBM, the manufacturers and universities worldwide, point to a sequence of next steps. Dates are each organisation's projections and slip often; the IRDS More Moore tables are a good public summary (IRDS, imec).
- Nanosheet refinements through the late 2020s: more and thinner sheets, better contacts, backside contacts to source and drain.
- Forksheets: n- and p-type stacks side by side with a thin dielectric wall between them, so they can sit closer.
- Complementary FETs (CFETs): the n-type stack built on top of the p-type stack (the last panel of the first figure), roughly halving the footprint of a CMOS pair. TSMC, Intel, Samsung and imec have shown working stacked devices and simple circuits at IEDM since 2023; separate contacts to the upper and lower devices are the hard part. Most projections put CFETs in production in the 2030s.
- 2D-material channels: sheets of MoS₂ or WSe₂ three atoms thick keep their mobility where silicon's collapses. The obstacles are growing defect-free monolayers on 300 mm wafers, low-resistance contacts and a good p-type material. A team at Fudan University reported a 32-bit RISC-V processor of about 5,900 MoS₂ transistors in 2025 (Ao et al.), a milestone for the material — with micrometre channels, running at kilohertz, which shows how far it is from a leading-edge product.
Beyond the transistor, much progress now comes from 3D integration (stacking dies with hybrid bonds at micrometre pitch) and design–technology co-optimisation, in which the cell library, wiring rules and process are designed together — topics of the next lesson.
Common misconceptions
Misconception
"Moore's law is dead." Transistor counts per chip and per package have kept growing, and transistor density still improves with each generation. What ended in the mid-2000s was Dennard scaling — the free speed and constant power density — and by most analyses the cost per transistor now falls much more slowly than it did before about 2012. Whether Moore's law is "dead" depends on which of these you mean.
Misconception
"Smaller transistors are always faster." Below about 20 nm, wire resistance rises steeply (thin copper lines suffer from electron scattering at their surfaces and grain boundaries, and the barrier liners take up a growing share of the cross-section), so a chip's speed is often limited by its wires, not its transistors. Newer processes improve speed at a given power by perhaps 10–15% per generation, not the 40% of the Dennard era.
Misconception
"Chips are made in one country." A leading-edge chip depends on lithography tools from the Netherlands with German optics, process tools from the United States and Japan, specialty chemicals and wafers largely from Japan, design software from the United States, fabrication mostly in Taiwan and Korea, and packaging and testing across East and Southeast Asia. No country has the whole chain.
Exercises
1. Transistor scaling
Warm-upHumanA designer shrinks a chip by per generation (so transistors per area double each time) but keeps the supply voltage fixed and raises the clock frequency by per generation, as the Dennard era had. Switching activity stays the same. By what factor has the power density grown after four generations?
2. Transistor scaling
CoreHumanA transistor's subthreshold swing is 75 mV per decade. If its threshold voltage is lowered by 150 mV (to make it faster), by what factor does its off-state leakage current increase?
3. Transistor scaling
Warm-upHumanWhat was the main physical reason that Dennard scaling ended in the mid-2000s?
4. Transistor scaling
CoreHumanWhich of these are genuine advantages of gate-all-around nanosheet transistors over FinFETs? Enter the letters.
- A. The gate surrounds the channel on all sides, giving better electrostatic control and lower leakage at short gate lengths.
- B. The sheet width can be drawn at any value, so drive strength is not limited to whole numbers of fins.
- C. They need no lithography for the channel.
- D. They have a fundamentally lower subthreshold swing than 60 mV per decade at room temperature.
- E. Several sheets stacked vertically give more drive current per unit of footprint.
5. Transistor scaling
CoreHumanUsing the Rayleigh criterion with , what is the smallest pitch (not half-pitch) that a single high-NA EUV exposure ( nm, NA = 0.55) can print? Give the answer to the nearest nanometre.
6. Transistor scaling
CoreComputationalImplement the lesson's cost model. dies_per_wafer(area) uses for a 300 mm wafer, rounded down to an integer; poisson_yield(area, d0) is with in mm² and in defects per cm²; and cost_per_good_die(wafer_cost, area, d0) divides the wafer cost by the expected number of good dies. Then set MONOLITHIC to the cost of one good 800 mm² die and CHIPLETS to the cost of four good 200 mm² dies (ignore packaging), both for a USD 25,000 wafer at per cm².
7. Transistor scaling
StretchComputationalWrite photons(dose_mj_cm2, wavelength_nm, side_nm), the expected number of photons landing on a square of resist side_nm on a side at the given dose, and dose_for_noise(fraction, wavelength_nm, side_nm), the dose in mJ/cm² at which the Poisson shot noise in that square equals fraction. Use with J s, m/s. (1 mJ/cm² = 10 J/m².)
8. Transistor scaling
Warm-upHumanWhich statement about a "2 nm" process such as TSMC N2 or Intel 18A is accurate?
Further reading and resources
- Cramming more components onto integrated circuits · Gordon E. Moore · 1965The original 1965 article, reprinted in the Proceedings of the IEEE in 1998.
- Design of ion-implanted MOSFET's with very small physical dimensions · R. H. Dennard, F. H. Gaensslen, H.-N. Yu, V. L. Rideout, E. Bassous, A. R. LeBlanc · 1974The paper that set out constant-field scaling.
- Dark silicon and the end of multicore scaling · H. Esmaeilzadeh, E. Blem, R. St. Amant, K. Sankaralingam, D. Burger · 2011
- FinFET — a self-aligned double-gate MOSFET scalable to 20 nm · D. Hisamoto et al. · 2000
- Stacked nanosheet gate-all-around transistor to enable scaling beyond FinFET · N. Loubet et al. · 2017
- International Roadmap for Devices and Systems (IRDS) · IEEEThe industry roadmap; its More Moore chapter tabulates pitches, device architectures and dates.
- TSMC 2nm technology · TSMC
- Intel 18A · Intel FoundryRibbonFET gate-all-around transistors and PowerVia backside power.
- Samsung Foundry · Samsung
- RapidusJapan's new 2 nm-class foundry, with its pilot line in Chitose, Hokkaido.
- EUV lithography systems · ASML
- imecThe Belgian research centre whose public roadmap and papers cover forksheets, CFETs and 2D channels.
- A RISC-V 32-bit microprocessor based on two-dimensional semiconductors · M. Ao et al. · 2025