Brian WilcoxZero → HeroSand → Compute: Silicon Fabrication

Fabrication · Physics · Supply Chain

From sand to compute.

Under every program is a slice of purified rock with billions of microscopic switches carved into it, flipped on and off by voltage. This is the full pipeline from sand to working chip: the intuition, the physics, the math, and the real path to building it yourself.

~100
Mask layers
per leading-edge chip
13.5 nm
EUV light
from a tin plasma
$20B+
Per fab
most capital-intense industry
9
Live simulators
runnable, not pictures
↓ scroll · a visual primer on how chips are made

01 / The mental model

A chip is a billion switches you printed with light

Before any process makes sense, you need the right abstraction: a modern chip is an enormous lattice of voltage-controlled switches (transistors), wired into logic. The entire job of a fab is to print those switches and wires, layer by layer, onto silicon.

// the mental model A transistor is the physical switch that carries out the 1s and 0s under every computation: the real hardware a logical decision runs on. Fabrication is how those switches get printed into a crystal, using ultraviolet light to lay down the pattern.

The transistor = a faucet for electrons

The MOSFET has three terminals: source, drain, and gate. Current wants to flow source→drain, but an insulated gate sits in the middle like a valve. Enough voltage on the gate pulls a conductive channel into existence: the faucet opens, electrons flow, the switch is ON (1). Drop the voltage and the channel vanishes, leaving it OFF (0).

mosfet.sim drag the gate voltage

The gate doesn't touch the current; it controls it through a thin insulator, purely by electric field. That insulator is why "gate oxide" is one of the most obsessed-over layers in the whole process.

Vth ≈ 0.40 V → output 0
A single transistor is ~10–50 nm today. A flagship chip packs 50–100 billion of them.

Why silicon, and what "doping" really does

Metals always conduct; glass never does. Silicon is a semiconductor, in between, and controllable, which is the whole point. Pure silicon barely conducts. You change it by doping: mixing in trace impurities. n-type (phosphorus) donates spare electrons; p-type (boron) creates "holes."

// analogy Pure silicon barely conducts. A few parts-per-million of dopant flips that behavior. Put n-type and p-type next to each other and you've built a diode: a one-way valve. Stack the right regions and you get a transistor.

The abstraction stack: from software down to switches

This is the map for the whole guide. Fabrication builds the bottom layers physically; everything above the transistor is software, everything below it is physics.

L7
Application software: the program you use
L6
Compiler / OS: code → instructions
L5
ISA, x86 / ARM / RISC-V: the contract
L4
Microarchitecture: pipelines, caches, cores
L3
Logic gates: AND / OR / NOT from transistors
L2
Transistors: the switches ← FAB BUILDS HERE
L1
Doped silicon + metal: electrons & physics

The rest of this guide walks L1 → L2: the physical manufacture of switches and wiring.

02 / Sand → wafer

Making a flawless canvas out of rock

You can't print circuits on dirty, lumpy material. Step one of any fab is producing a single crystal of silicon, 99.9999999% pure (nine nines), sliced into mirror-flat discs called wafers. Every later step happens on top of this.

wafer.render
A finished wafer: one continuous crystal, polished to a mirror, tiled with a grid of identical dies. The iridescent sheen is real thin-film interference.
the_journey_of_a_silicon_atom
quartz sandSiO₂ metallurgical98% Si polysilicon9N pure crystal ingotCzochralski wafer300 mm disc dies printed
Quartz → crude silicon → polysilicon → one giant single crystal → sliced & polished into wafers.

Czochralski growth: pulling a perfect crystal

Molten silicon is held just above melting. A tiny seed crystal touches the surface and is slowly rotated and pulled upward. Atoms latch onto the seed's lattice in perfect alignment, so the whole boule (often 2 m long, 300 mm across) is one continuous crystal with no grain boundaries.

// analogy A single crystal is a flawless, repeating lattice: every atom in its assigned place. Any misalignment is a defect that scatters electrons and ruins device behavior. Czochralski exists to get the exact same atomic arrangement everywhere, with no grain boundaries to interrupt it.
// reality check A 300 mm wafer costs a few hundred dollars blank. After ~3 months and ~1,000 process steps it can be worth $15,000–$30,000+. The value is entirely in the printing.

03 / Photolithography

The printing press at the heart of everything

Lithography turns a circuit drawing into a physical pattern. It's the single most important (and most repeated) step. A leading chip runs this loop 50 to 100 times, once per layer. Master this and you understand the fab.

the_core_insight.md
# Lithography prints one layer, then repeats

mask = the source pattern (the layout file)
light = projects that pattern onto the wafer
resist = the coating the pattern develops into
etch = removes material through the openings

# then run the whole sequence again for the next layer:
# each layer stacks on the one below it, so the chip
# is built up physically, layer over layer.
litho.step one pass = one layer

A side-view cross-section of the wafer surface. Step through the loop.

1. Oxidized wafer
Start with silicon and grow a layer of insulating oxide on top.

Resist: the light-sensitive paint

The wafer is spin-coated with photoresist, a polymer whose chemistry changes where light hits it. Positive resist becomes soluble where exposed (the lit pattern washes away); negative does the opposite. After developing, you're left with a stencil protecting some areas and exposing others.

// why fab photos are always yellow Lithography bays use amber light on purpose. Photoresist reacts to short-wavelength (blue/UV) light, so white lighting would expose it accidentally. Yellow light has none of those wavelengths. That amber glow is the actual color of the most important room in the building.

The stepper: a reduction camera

The exposure tool shines light through the mask, then through a lens that shrinks the image 4× onto the wafer, and "steps" across to repeat it per die. Reduction is a gift: a mask with 4×-larger features is far easier to make. The lens does the hard part.

04 / Resolution & EUV

The math of small, and the machine that prints it

How tiny can you print? It comes down to one equation from optics. Understanding it explains the entire 40-year race down to today's nodes, and why a single Dutch company controls the future.

rayleigh.py
# smallest printable feature (critical dimension)
CD = k₁ · λ / NA

λ = wavelength of light
NA = numerical aperture (how wide the lens grabs light)
k₁ = process cleverness factor (theoretical floor ≈ 0.25)

# to print smaller: shrink λ, raise NA, or lower k₁.
# the whole industry is a 40-year war on these 3 numbers.
resolution.explore fight the equation

Tap a preset to jump to a real technology generation, or drag the sliders. Watch how many lines you can pack into the same width.

g-line 436nm i-line 365nm KrF 248nm ArF 193nm ArF immersion EUV 13.5nm High-NA EUV
min feature (CD)16 nm
vs a human hair5000× smaller
lines in 1 µm31
EUV's leap from 193 nm to 13.5 nm is a 14× shorter wavelength: the biggest single jump in litho history, and why sub-7 nm chips exist.

Why EUV is borderline impossible

To get 13.5 nm light you can't use a normal lamp or laser. Here's what ASML's machines actually do, 50,000 times per second:

euv_source.sim laser-produced tin plasma

One droplet-to-light cycle, slowed roughly a million times. Play the loop or step it phase by phase. In a real machine this whole sequence repeats 50,000 times every second, a fresh tin droplet vaporized every 20 microseconds.

1. Tin droplet falls
phase1/5
droplet rate (real)50,000 /s
this loop slowed~106×
A ~25 µm droplet of molten tin is fired across the vacuum chamber toward the plasma point.
step 1

Vaporize tin

A droplet of molten tin is shot through a vacuum. A laser pre-pulse flattens it into a pancake.

step 2

Blast it twice

A massive CO₂ laser hits the flattened droplet, making a ~500,000 °C plasma that emits 13.5 nm light.

step 3

Mirrors, never lenses

EUV is absorbed by everything: glass, air, your hand. Optics are ultra-smooth Mo/Si multilayer mirrors, all in vacuum.

step 4

Collect the scraps

Only a few percent of light survives to the wafer. The machine weighs ~180 tons, costs ~$200M, ships in 40+ crates.

// the chokepoint ASML (Netherlands) is the only company on Earth that makes EUV machines. No EUV, no leading-edge chips. This single dependency is why semiconductors are now geopolitical (see §09).

When you can't shrink λ: multi-patterning & high-NA

From 2007–2018 the industry was stuck at 193 nm yet kept shrinking, via multi-patterning: split one dense pattern into 2–4 masks/exposures that interleave on the wafer (expensive, like rendering one frame in multiple passes). EUV collapsed many passes back into one. Next up: High-NA EUV (NA 0.55), rolling out now.

05 / Deposit · etch · dope

Adding, removing, and re-flavoring material

Lithography only makes the stencil. Three families of process actually change the wafer: deposition adds material, etching removes it, and doping changes the silicon's electrical personality.

Deposition: thin films, sometimes one atom at a time

MethodHowAnalogy
CVDGases react and deposit a solid filmSpray-paint a coat
PVDSputter/evaporate metal atoms onto the surfaceVapor-blast a layer on
ALDOne atomic layer per cycle, self-limitingOne exact layer at a time
EpitaxyGrow new crystal aligned to the wafer belowExtend the existing structure
// why ALD matters Atomic Layer Deposition is self-limiting: each cycle deposits exactly one molecular layer and then stops, no matter how long you wait. That is the only way to build the angstrom-thin, perfectly uniform gate insulators modern transistors need.
ald_cycle.sim one atomic layer per cycle

Four self-limiting sub-steps make one cycle. Precursor A saturates every open surface site and then stops (extra molecules bounce off, that is what "self-limiting" means). Purge, then co-reactant B converts that layer into film and frees the sites again. Play it or step through, and watch the film grow exactly one layer per cycle.

A · precursor pulse
cycles complete0
film thickness0.0 Å
layers deposited0
Each cycle adds ~1 Å (one molecular layer). Thickness is set by counting cycles, not by timing a flow, which is why ALD hits angstrom precision no other method can match.

Etching: anisotropic is everything

Wet etching dunks the wafer in chemicals: fast, but eats sideways too (isotropic), blurring features. Dry / plasma etching fires energized ions straight down, carving sharp vertical walls. At nanometer scale you need walls, not slopes.

etch_profile.compare
WET (isotropic) rounded, undercut DRY / PLASMA (anisotropic) straight vertical walls
Plasma ions travel in straight lines, so they cut down without cutting sideways, which is essential below ~100 nm.

Doping: the random walk hiding inside Fick's law

To make n-type or p-type regions, drive dopant atoms into the silicon. Ion implantation literally accelerates dopant ions in a particle accelerator and fires them into the wafer; a high-temperature anneal repairs the lattice and lets atoms settle. How they spread is pure statistics.

diffusion.py
# Fick's law looks like calculus, but the intuition is a coin flip.
# Each dopant does a random walk: hop up/down/left/right, no memory.
# Aggregate millions of walkers → a smooth bell-shaped profile emerges.

∂C/∂t = D · ∂²C/∂x²

# that's the entire content of the diffusion equation.
fick_random_walk.sim

Every dot is one dopant atom doing a random walk into the silicon. The bars on the right are the concentration-vs-depth profile building up in real time: Fick's law happening atom by atom.

Higher temperature → bigger random jumps → deeper, faster penetration. No atom "knows" where to go; the smooth profile is an emergent average.

CMP: polish flat, then do it all again

After stacking layers, the surface gets bumpy. Chemical-Mechanical Planarization grinds it back to a mirror flat, because the next lithography step needs a level surface to stay in focus (recall the tiny depth-of-field at 13.5 nm).

// analogy CMP is the cleanup pass between layers. It grinds the bumpy surface back to a mirror so the next lithography step lands in focus. Skip it and every later layer inherits the bumps, each one compounding on the last.

06 / Transistors & wiring

Build the switches first, then wire the city

A fab's flow splits in two. FEOL (front-end-of-line) builds the transistors in the silicon. BEOL (back-end-of-line) builds the 10–15+ stacked layers of copper wiring connecting billions of them: a microscopic multi-level highway system.

FEOL: the switches

Front-end-of-line

Wells, gates, gate oxide, source/drain doping. Where the actual transistors form. Highest precision, smallest features.

BEOL: the wiring

Back-end-of-line

Copper "damascene" interconnect in many stacked layers, separated by low-k insulator, joined by vertical vias. Lower layers dense and fine; upper layers fat power/clock rails.

Transistor evolution: chasing control of the channel

As transistors shrank, the flat gate lost grip and current leaked even when "off" (wasted power, heat). The fix each generation: wrap the gate around more of the channel.

gate_geometry.evolution end-on cross-section

Looking down the channel (source→drain runs into the screen). Toggle the generation: watch the orange gate wrap around more of the blue channel, and the off-state leakage current bleeding through drop as electrostatic control tightens.

Planar FinFET ~2011 GAA nanosheet ~2022+
planar MOSFET
gate wraps1 side
channel coverage25%
off-state leakagehigh
Planar (1 side) → FinFET (3 sides) → Gate-All-Around nanosheets (fully surrounded). More wrap = tighter grip on the channel = exponentially less sub-threshold leakage at the same tiny size. Leakage electrons here follow that coverage.
// "3nm" is marketing, not a measurement Node names like 5nm, 3nm, 2nm stopped matching any physical dimension around 2010. Nothing on a "3nm" chip is actually 3 nm: it's a brand for a generation's density & performance. The real metric engineers use is transistor density (millions per mm²).

Copper interconnect: the damascene trick

Copper is hard to etch, so wiring is built in reverse: etch trenches in the insulator, fill with copper, then CMP away the excess so only wires-in-trenches remain. Named after the inlay metalwork of Damascus. Repeat per metal layer.

07 / Test & package

Sort the survivors, then armor them

A finished wafer is a grid of identical dies, but they aren't all equal; defects are scattered randomly. The back end is about finding the good ones, cutting them out, and connecting them to the world.

Probe, then bin

Before dicing, every die is electrically tested on the wafer ("wafer probe"). Working dies are binned by speed and power efficiency.

// why your CPU has speed grades Binning is A/B sorting at the hardware level. The same design comes off the line at a spread of qualities. Best-clocking dies become the premium part; weaker ones sell at lower clocks; partly-defective ones may have a bad core disabled and ship cheaper. One wafer → a whole product ladder.

Dice, package, bond

The wafer is sawn or laser-cut into individual dies. Each good die is mounted in a package that protects it and fans its microscopic pads out to pins/balls you can solder to a board.

Advanced packaging: where the action moved

With single-chip shrinking getting brutal, gains now come from combining chips: chiplets, 3D stacking, silicon interposers (TSMC's CoWoS), and stacked memory (HBM). Modern AI accelerators are really several dies fused into one package.

why_chiplets.md
# The chiplet payoff is yield math (see §08):
# a defect kills a whole die, so smaller dies waste
# less silicon per defect. Four small dies survive
# defects far better than one big one, and you can
# even mix process nodes. Economics force the packaging.
packaging.compare monolithic vs chiplets

Same total logic, two ways to build it. Toggle between one big monolithic die and four chiplets sitting on a silicon interposer next to HBM memory stacks. Drag the die size up and watch the yield gap explode: the chiplet approach uses the exact same e^(−A·D) model from §08, just on four small dies instead of one giant one.

Monolithic die Chiplets + interposer + HBM
good-silicon yield-
monolithic (1 die)-
chiplets (4 dies)-
good silicon rescued-
A defect kills whatever die it lands in. Splitting one A-area die into four A/4 dies means each defect scraps a quarter of the silicon, not all of it, so good-die yield jumps from e^(−A·D) to e^(−A·D/4). That gap is the entire economic case for chiplets, CoWoS, and HBM.

08 / Yield & economics

The equation behind chip economics

Defects are random and unavoidable. The fraction of working dies is the yield, and it dictates cost, chip size, and why chiplets exist. It's also simple enough to write in five lines of code.

yield.py
# Poisson yield model: defects land randomly at density D.
# chance a die of area A catches ZERO defects:

Yield = e^(−A·D)

# double the die area → yield falls off a cliff, exponentially.
yield.sim defects vs die size

Random red defects scatter on a 300 mm wafer. Any die containing a defect dies (dark). Shrink the die or clean up the process and watch yield climb.

good dies-
dead dies-
yield (simulated)-
yield (e^−AD)-
A giant GPU die (~26 mm) at the same defect density yields a fraction of what a small die does. That's the entire business case for chiplets.

Why a fab costs more than an aircraft carrier

capital

$10B–$20B+

One leading-edge fab. EUV tools alone run ~$200M each, and you need dozens.

cleanliness

Class 1 cleanroom

Air ~10,000× cleaner than a hospital OR. A single dust speck destroys a die.

scale

~3 leading-edge players

Only TSMC, Samsung, and Intel do the bleeding edge. The cost barrier created a near-monopoly.

learning curve

Yield ramps for years

New nodes launch at low yield and improve through relentless defect-hunting. Early yield is the moat.

09 / The supply chain

No single country can build a chip alone

A finished phone chip might be designed in California on British IP, using American software, printed in Taiwan on a Dutch machine, with Japanese chemicals and German optics. The supply chain is a globe-spanning relay with brutal chokepoints.

Design tools (EDA)
Synopsys · Cadence · Siemens EDA
IP cores ⚠
Arm (UK) · RISC-V (open)
↓ the blueprint
Fabless designers
Nvidia · Apple · AMD · Qualcomm · Broadcom
↓ send layout to be printed
Leading-edge foundry ⚠
TSMC (Taiwan, ~90% of leading edge) · Samsung · Intel
↓ which depend on…
Litho machines ⚠⚠
ASML (NL): sole EUV maker
Other equipment
Applied Materials · Lam · KLA · Tokyo Electron
Materials ⚠
Shin-Etsu/SUMCO wafers · JSR/TOK resist (Japan)
↓ then back end
Packaging (OSAT)
ASE · Amkor · JCET · TSMC CoWoS
Stacked memory
SK Hynix · Samsung · Micron (HBM)
// the chokepoints (⚠ above) Each red node is a single point of failure the world depends on: ASML for EUV, TSMC/Taiwan for printing leading-edge chips, Japan for irreplaceable materials, Arm + US EDA for designs. This is why export controls, the US CHIPS Act (~$52B), and Taiwan's "silicon shield" are front-page geopolitics.

Three business models, one ecosystem

ModelDoes whatExamples
FablessDesigns chips, owns no factoryNvidia, Apple, AMD, Qualcomm
FoundryManufactures others' designsTSMC, GlobalFoundries
IDMDesigns and manufacturesIntel, Samsung, Micron
// analogy It mirrors cloud computing. Fabless = design the chip and let someone else run the factory. Foundry = AWS for silicon: TSMC is the shared plant everyone rents. IDM = owns its factories end to end. That design-versus-manufacture split is what let chip startups exist without spending $20B on a fab.

10 / The build path

What "from scratch" actually means for you

Honest truth: you can't build a leading-edge chip in a garage: that needs ~$20B and a Dutch monopoly's blessing. But "0 → hero" is real, and there are three legitimate ladders you can start climbing this week, in order of accessibility.

ladder 1: software, fully real

Design & tape out a real chip

Write hardware in Verilog / Chisel, target the open SkyWater SKY130 PDK, run the open OpenLane / OpenROAD flow (synth → place → route → GDSII), and submit to a shared shuttle like Tiny Tapeout to get a physical chip back in the mail. You'll touch every concept here as a real tool, not a metaphor.

ladder 2: simulate the physics

Model every step in code

The nine simulators in this guide are runnable models: extend them. Build the Rayleigh tool, a Monte-Carlo diffuser, a Poisson yield estimator, a logic-gate sim from MOSFET equations. This is where a software brain has a genuine edge.

ladder 3: physical, hobbyist

Make transistors at home

Sam Zeloof built working ICs in a home lab (~1960s–70s nodes); Jeri Ellsworth made hand-built transistors. Not nanometers, but real lithography, etching, and doping at micron-ish scale with serious safety discipline. Proof the whole loop is comprehensible by one person.

foundations to shore up

The prerequisites

Digital logic & boolean algebra → a little semiconductor physics (band gap, P-N junction, MOSFET I–V) → an HDL → the open EDA flow. Each maps onto a section above: §01 logic, §05 physics, §06 devices, §10 tooling.

The 0 → hero sequence, one line each

wk 1
Internalize §01: transistor = switch, doping = trace impurities, the abstraction stack
wk 2
Build a logic gate from boolean primitives, then learn an HDL
wk 3
Walk the litho loop (§03) & the resolution math (§04) until intuitive
wk 4
Run OpenLane on SKY130 with a tiny design → produce a GDSII layout
wk 5+
Submit to a Tiny Tapeout shuttle · meanwhile model yield & diffusion in code
hero
You hold a chip you designed, and you understand every step that built it
// the mental shift The hardest part isn't the physics: it's seeing the whole thing as one repeatable pipeline. A mask defines the pattern, machines print and test it layer by layer, and the working chips are sorted by speed. Once you see it as a pipeline, the industry stops looking like magic and starts looking like engineering.

More from Zero → Hero

Browse all 15 guides →