Semiconductor Manufacturing
Transistors to Processors · Hardware Foundations track
7 min read
In this module
How Chips Are Made
Every processor starts as a cylinder of ultra-pure silicon. Getting from there to a finished chip with billions of transistors is one of the most intricate manufacturing processes humans have ever devised. Understanding it explains why specific process nodes exist, why chips cost what they do, and why transistor scaling matters for security — more transistors means more room for hardware security features that were simply impossible at coarser geometries.
From Sand to Wafer
Silicon is extracted from silicon dioxide — sand — and refined to 99.999999999% purity (eleven nines). That ultra-pure silicon is melted, and a seed crystal is slowly drawn upward through the melt using the Czochralski process, growing a single-crystal ingot. The ingot is sliced into wafers — 300 mm diameter in modern fabs — polished to near-atomic flatness, and sent to the fab.
Each wafer eventually yields hundreds of dies. Larger dies mean fewer per wafer and a greater chance that any given die intersects a manufacturing defect.
Yield is the percentage of functional dies per wafer. A single particle of dust can destroy a transistor. Modern fabs operate in cleanrooms thousands of times cleaner than a hospital operating theatre, yet yield is still a critical economic factor. Larger dies have lower yield because each die has more area exposed to potential defects.
Photolithography
The core of chip manufacturing is photolithography - using light to transfer circuit patterns onto the silicon wafer.
Each layer of the chip requires a separate mask and a full cycle of coat, expose, develop, etch, and deposit. A modern processor has 80-100 mask layers. The entire process takes 2-3 months from bare wafer to packaged chip.
EUV Lithography
For decades, fabrication used deep ultraviolet (DUV) light at 193 nm wavelength. But as features shrank below 7 nm, even with tricks like multi-patterning, DUV could not reliably print such small features.
Extreme ultraviolet (EUV) lithography uses 13.5 nm wavelength light, enabling direct patterning of features below 7 nm. EUV machines are among the most complex devices ever built:
- The light source works by firing a laser at tiny tin droplets, creating a plasma that emits EUV light
- EUV light is absorbed by glass, so the optics use multilayer mirrors instead of lenses
- The entire light path must be in a vacuum
- Each machine costs approximately $200 million and weighs over 180 tonnes
EUV is essential for 5 nm, 3 nm, and smaller process nodes. Only one company, ASML, manufactures EUV lithography systems.
Transistor Architectures
As transistors shrink, the physics of controlling current flow becomes increasingly challenging. The industry has evolved through several transistor architectures:
Planar Transistors
The original MOSFET design: a flat gate electrode sits above the channel between source and drain. This worked well down to approximately the 22 nm node. Below that, the gate could not adequately control the thin channel, leading to excessive leakage current (current flowing even when the transistor is supposed to be off).
FinFET (3D Transistors)
Introduced at the 22 nm node by Intel in 2012, the FinFET (Fin Field-Effect Transistor) wraps the gate around a vertical fin of silicon on three sides. This tri-gate structure gives the gate much better control over the channel:
FinFETs cut leakage sharply while improving switching speed. They have been the dominant transistor architecture from 22 nm through 5 nm.
Gate-All-Around (GAA) and Nanosheets
At the 3 nm node and below, even FinFETs struggle with leakage. Gate-All-Around transistors (called nanosheets by Samsung and TSMC, and RibbonFET by Intel) stack horizontal sheets of silicon with the gate material completely surrounding each sheet. This provides the ultimate electrostatic control.
Samsung was first to manufacture GAA transistors, at its 3 nm node (3GAE) in 2022. Intel had planned to debut RibbonFET on its 20A node but cancelled 20A in September 2024; RibbonFET instead arrived with Intel 18A, whose first products (Panther Lake) entered volume production in late 2025. TSMC moved to nanosheets with its N2 (2 nm) process, which also entered volume production in late 2025.
Process Nodes and Naming
A process node was originally named after the smallest feature size in nanometres, roughly the gate length. That link loosened long ago: Intel's 90 nm process (2003) already used transistors with a 50 nm gate. Since approximately 2010, node names have become marketing labels that no longer correspond to any physical dimension.
| Node Name | Year | Approx. Gate Length | Transistor Type |
|---|---|---|---|
| 90 nm | 2004 | ~50 nm | Planar |
| 45 nm | 2007 | ~35 nm | Planar |
| 22 nm | 2012 | ~22 nm | FinFET (Intel) |
| 14 nm | 2014 | ~20 nm | FinFET |
| 7 nm | 2018 | ~10 nm | FinFET |
| 5 nm | 2020 | ~7 nm | FinFET |
| 3 nm | 2022 | ~5 nm | FinFET/GAA |
The gate lengths are rough and vary between foundries and between transistor types on the same chip, so read them as orders of magnitude. Different foundries use different conventions. TSMC's "3 nm" and Samsung's "3 nm" have different densities and characteristics — the node label means nothing without context. The number that actually matters is transistor density (millions of transistors per mm²).
Process node names like '5 nm' or '3 nm' do not describe any physical dimension on the chip. They are marketing terms used by foundries. Actual transistor gate lengths at '5 nm' are closer to 7 nm. The useful comparisons are transistor density, power consumption, and performance, not the name.
Dennard Scaling and Its End
For decades, as transistors shrank, they also became faster and used less power per transistor, following Dennard Scaling (1974). The key insight: if you reduce all dimensions by a factor of k, voltage drops by k, current drops by k, and power per transistor drops by k squared, while frequency can increase by k.
This broke down around 2005-2006. Below roughly 65 nm, leakage current became significant enough that voltage could no longer be reduced proportionally. This ended Dennard scaling and triggered the shift from ever-faster single cores to multi-core processors.
The consequences are fundamental:
- Clock speeds have plateaued at roughly 4-5 GHz for over a decade
- Performance gains now come from more cores and wider execution engines
- Power consumption is the primary design constraint
- Dark silicon - portions of the chip that cannot be powered simultaneously due to thermal limits - is a real engineering challenge
Security Implications of Manufacturing
Manufacturing has direct security consequences:
- Hardware trojans: malicious modifications inserted during fabrication. Modern chips are manufactured by third-party foundries (TSMC, Samsung, GlobalFoundries) that the design owner doesn't control. Finding a trojan among billions of transistors is genuinely hard.
- Side channels from manufacturing variation: no two transistors are identical. Process variation creates measurable differences in timing, power draw, and EM emissions — all of which can bleed secret data.
- Physically Unclonable Functions (PUFs): that same variation becomes a security feature. PUFs harvest the unique, random physical characteristics of each chip — SRAM power-up states, ring oscillator frequencies — to generate per-chip keys that are impossible to clone precisely because no two chips are physically identical.
- Supply chain integrity: verifying that chips haven't been tampered with, counterfeited, or swapped out requires everything from X-ray inspection to carefully designed functional tests.