What Is a Node in Semiconductor Manufacturing?

A node is a generation of chip-manufacturing technology: the set of methods a factory uses to build a chip, labeled with a number such as 5nm or 2nm. Each new node packs more and better transistors into the same space, which is why phones and laptops advertise it.

But the number in the name no longer measures anything you could find on the chip. It has become closer to a brand name for a generation.1Tessolve. 7nm vs 5nm vs 2nm: How to Choose the Right Semiconductor Process Node Within one company, a smaller number means newer, though not a precise size, and two companies’ nodes with the same number may not be equal.

Where the Nanometer Number Came From

A computer chip is built from billions of transistors, tiny electrical switches that turn on and off to process information. Each transistor has a gate, which acts like a valve that opens and closes the flow of electrons.

The gate sits between two electrical contacts, the source and the drain, and gate length is the distance between them that the gate spans.2IEEE Spectrum. A Better Way to Measure Progress in Semiconductors2IEEE Spectrum. A Better Way to Measure Progress in Semiconductors A shorter valve can open and shut faster, so a shorter gate generally meant a faster-switching transistor. That made it the most important size for performance.2IEEE Spectrum. A Better Way to Measure Progress in Semiconductors Shrinking every transistor also let more of them fit on a chip, which was the real prize.

For a long time the gate length was about the same as the metal half-pitch, which is half the spacing between neighboring wires on the chip.2IEEE Spectrum. A Better Way to Measure Progress in Semiconductors That shared size became the node number, as in names like the “90 nm process.”3Renesas. Semiconductor process technology; History, trends and evolution A nanometer is a billionth of a meter. Even older generations were measured in microns, a thousand times larger than nanometers.3Renesas. Semiconductor process technology; History, trends and evolution

Because gate length and wire spacing shrank together, one figure let the whole industry describe a generation. Each generation was roughly a 0.7 times shrink in linear size, which about halved the area each transistor needed, so the same chip area could hold about twice as many transistors. That regular rhythm is the pattern often called Moore’s law, and node names were the labels on its steps.

Why Today’s Node Names Are Mostly Marketing

Since about 1997, node names have not matched any actual dimension on the chip, and have been off by almost a factor of 2 in both directions.4EEJournal. No More Nanometers Around 2011 the gap grew as names shifted to smaller numbers than the real gate length.5Wikipedia. 5 nm process

The reason is partly geometry. On a flat transistor, gate length is one clear measurement you can point to. On a fin-shaped transistor, the gate wraps around a tall, thin fin, so the length along the fin, the fin’s width and its height all matter, and no single one captures the design. Chipmakers had no obvious measurement to keep naming nodes after, and the names kept following the old pattern of shrinking numbers. Names like 7nm or 5nm now signal relative progress and efficiency gains, not an exact gate length.1Tessolve. 7nm vs 5nm vs 2nm: How to Choose the Right Semiconductor Process Node

Commercial motives play a part too. Some chipmakers promote node names to show they lead the race, and many insiders call the numbers marketing terms.6Semiconductor Engineering. Nodes Vs. Nodelets The number is useful for ordering one company’s own generations, but weak for comparing companies. TSMC has even moved to names like A16, dropping the nanometer label for its next generation.7TSMC. What Is a Node in Semiconductor Manufacturing? Real-world speed and battery life depend on the chip design as well as the node, which is why reviews and benchmark results tell you more than the label.

Better Ways to Compare Nodes

The most common honest yardstick is transistor density. Intel proposed measuring generations by the number of transistors per square millimeter, using standard logic building blocks.8IEEE Spectrum. Intel Now Packs 100 Million Transistors in Each Square Millimeter By that measure Intel reached about 100 million transistors per square millimeter on one generation.8IEEE Spectrum. Intel Now Packs 100 Million Transistors in Each Square Millimeter

Among current 2nm-class processes, TSMC N2 reaches about 313 million per square millimeter, Intel 18A about 238 million and Samsung SF2 about 231 million.9UST. TSMC N2 vs Samsung 2GAP vs Intel 18A: The 2nm Foundry Decision Processes with similar names can differ by more than a third in density.

The phrase “standard logic building blocks” matters because density depends on which part of a chip you measure. Memory can be packed more tightly than logic, the circuits that make decisions. Intel’s proposal measures typical logic building blocks so that different processes are judged on a similar footing. It is still a simplification, because a real chip mixes many kinds of circuits.

Density alone isn’t the whole story either. Two chips with equal density can differ in speed and power use, so makers quote those gains between generations alongside density. TSMC, for example, quotes A16 as 8-10% faster and 15-20% lower power than N2P. Treat all such figures as maker claims for comparable designs.

The IEEE roadmap takes another approach, proposing names like GxxMxxTx. These give the gate spacing and tightest wire spacing in nanometers, plus the number of stacked transistor layers.10IEEE International Roadmap for Devices and Systems. System and High-Volume-Manufacturing Driven More Moore Scaling Roadmap Density numbers are published mainly by industry analysts, not in phone ads, so independent tests of speed and battery life are the practical comparison.

How Transistors Changed Shape to Keep Shrinking

As flat transistors shrank, they leaked current when switched off and became harder to control. A transistor is supposed to block current when off, but in a very short flat transistor the gate controls the channel only from the top, so some current slips through underneath. Leaked current wastes battery and produces heat, which limits how many transistors can be packed and run together.

The fix was multi-gate designs, where the gate wraps several sides of the channel.11PubMed Central. Comprehensive Review of FinFET Technology: History, Structure, Challenges, Innovations, and Emerging Sensing Applications Wrapping the gate around more of the channel gives better control, higher speed and less leakage.11PubMed Central. Comprehensive Review of FinFET Technology: History, Structure, Challenges, Innovations, and Emerging Sensing Applications

Intel’s tri-gate fin design combined this 3D shape with new gate materials and strained silicon.122006 Symposium on VLSI Technology, 2006. Digest of Technical Papers.. Tri-Gate Transistor Architecture with High-k Gate Dielectrics, Metal Gates and Strain Engineering The fin-shaped transistor, known as a FinFET, became the industry’s workhorse for many generations.13imec. Entering the nanosheet transistor era

A flat transistor’s gate sits only on top of the channel, like a lid. A fin stands the channel up so the gate wraps three sides, like a saddle. Nanosheets stack thin ribbons with the gate surrounding each one completely.

From 2022 or 2023, leading chipmakers began moving from fin-shaped transistors to nanosheet designs for 3nm and 2nm chips.13imec. Entering the nanosheet transistor era Next comes CFET, which stacks the two types of transistors chips need on top of each other.14imec. CFET (complementary FET) Intel, Samsung and TSMC have all demonstrated CFETs, which could nearly double transistor density.15IEEE Spectrum. Intel, Samsung, and TSMC Demo 3D-Stacked Transistors Progress now comes partly from building upward, from 2D to 3D.16PubMed. From Planar to 3D: The Evolution of Complementary Field-Effect Transistor for Next-Generation Semiconductor Nodes

Printing Features Smaller Than Light Allows

Chip patterns are printed by shining light through a pattern onto a light-sensitive coating on a silicon wafer. The coating changes chemically where light hits it. After exposure, the wafer is washed so that either the lit or the unlit parts dissolve, leaving a pattern. That pattern then guides etching or the deposit of materials into the silicon, and a chip repeats this cycle for many layers.

Shorter wavelengths can draw smaller shapes, like a fine pen versus a thick marker. Yet 193-nanometer light is many times longer than today’s smallest features. Factories pushed it beyond its natural limit by printing a layer in several passes.17Advanced Etch Technology and Process Integration for Nanopatterning X. Advanced multiple patterning technologies for high density hexagonal hole arrays But each extra pass adds steps, cost and chances for misalignment.17Advanced Etch Technology and Process Integration for Nanopatterning X. Advanced multiple patterning technologies for high density hexagonal hole arrays

EUV lithography uses extreme ultraviolet light with a wavelength of 13.5 nanometers.18ZEISS SMT. High-NA-EUV Lithography: the next EUV generation That light is absorbed by air and glass, so the machines use mirrors in a vacuum.

New high-NA EUV machines target features as fine as 8 nanometers half-pitch.19Journal of Micro/Nanopatterning, Materials, and Metrology. Next step in Moore’s law: high NA EUV system overview and first imaging and overlay performance Reaching this required redesigned mirror optics that stretch the image in one direction.20SPIE Proceedings. Anamorphic high-NA EUV lithography optics

The Wiring Problem Inside Every New Node

Copper has long been used for chip wiring because it conducts electricity well.21PubMed Central. Graphene-All-Around Cobalt Interconnect with a Back-End-of-Line Compatible Process But copper’s resistance rises sharply once wires shrink below about 30 nanometers.21PubMed Central. Graphene-All-Around Cobalt Interconnect with a Back-End-of-Line Compatible Process In very thin wires, electrons scatter off the wire’s surfaces and grain boundaries more often, which slows their flow. Copper also needs a thin barrier lining to stop it spreading into surrounding material, and that lining takes up room that would otherwise carry current. As the wire narrows, the lining takes a bigger fraction of its width, so the conducting part shrinks faster than the wire does.

Higher current density in thin wires also speeds up electromigration, where moving electrons gradually wear wires out.21PubMed Central. Graphene-All-Around Cobalt Interconnect with a Back-End-of-Line Compatible Process The result is that overall signal delay is now often dominated by the wiring rather than transistor switching speed.22PubMed. Addressing interconnect challenges for enhanced computing performance Researchers are testing metals such as ruthenium and cobalt as replacements.23imec magazine. imec magazine September 2017 – How to solve the back-end-of-line RC delay problem?

Chips have two kinds of wiring. Signal wires carry information between transistors, and power wires feed electricity to every transistor. Putting both on the same side crowds them together like traffic on one road.

Backside power delivery moves the power wiring to the underside of the silicon wafer, separating it from signal wiring.24imec. Backside power delivery Power reaches transistors through wider, lower-resistance lines.24imec. Backside power delivery That frees room on the front and steadies the supply. In one 3nm-node design study, this cut voltage loss from over 10% of the supply to under 70 millivolts.25IEEE Xplore. A Comparative Study on Front-Side, Buried and Back-Side Power Rail Topologies in 3nm Technology Node

Why New Nodes Cost Billions

In the late 1960s a cutting-edge chip factory cost a few million dollars.26International Center for Law & Economics. From Moore’s Law to Market Rivalry: The Economic Forces That Shape the Semiconductor Manufacturing Industry By the early 2000s it cost low billions.26International Center for Law & Economics. From Moore’s Law to Market Rivalry: The Economic Forces That Shape the Semiconductor Manufacturing Industry Today a leading-edge US factory for nodes under 4nm can cost $10-20 billion or more.26International Center for Law & Economics. From Moore’s Law to Market Rivalry: The Economic Forces That Shape the Semiconductor Manufacturing Industry The cost reflects the precision needed near atomic scales, not just inflation.26International Center for Law & Economics. From Moore’s Law to Market Rivalry: The Economic Forces That Shape the Semiconductor Manufacturing Industry Only a few players, mainly TSMC, Samsung and Intel, can compete.27Microchip USA. The Future of Semiconductor Miniaturization

Much of that money goes to equipment, and EUV machines are among the most expensive tools a factory buys. A new node also needs entirely new process steps, so a factory often must be fitted out again for each generation. Clean rooms and the systems that keep them pure add further cost, since a single speck of dust can ruin chips.

Yield is the share of good chips from a wafer, and it strongly drives manufacturing cost.28IEEE Transactions on Semiconductor Manufacturing. Yield Learning Curve Models in Semiconductor Manufacturing A new node usually starts with low yields that must be raised quickly to make a profit.28IEEE Transactions on Semiconductor Manufacturing. Yield Learning Curve Models in Semiconductor Manufacturing

These costs reach consumers in a predictable way. Newest-node chips usually appear first in flagship phones and high-end computers. Cheaper devices often use older nodes that are already paid off, and older nodes stay widely used for cars, appliances and simpler chips.

How Small Can Nodes Get

At very small sizes, electrons can tunnel through barriers they normally couldn’t cross, making current hard to control.29npj Computational Materials. Atomic-level simulations predict transistor scaling limits An off switch is then not fully off, which wastes power and makes errors more likely. Because this is a feature of quantum physics, better engineering can reduce it but not remove it.

Tunneling also blocks further thinning of the gate’s insulating layer.30Science. Pushing the Limits At this scale, small differences in the number of added impurity atoms can change how a transistor behaves.30Science. Pushing the Limits At room temperature, transistors of any material face unacceptable errors around 5-nanometer gate lengths.31Applied Physics Letters. The fundamental downscaling limit of field effect transistors A “2nm” label does not mean chips have passed that limit, since the label isn’t a gate length.

Packaging offers another route. Chipmakers increasingly combine chiplets, smaller chips made on different nodes, inside one package.32IMAPSource Proceedings. Market & Technology Trends for the Fanout and 2.5D/3D Packaging Technology This mixes newest and older nodes as transistor shrinking grows costly.32IMAPSource Proceedings. Market & Technology Trends for the Fanout and 2.5D/3D Packaging Technology Chiplet packaging is also generally easier to manufacture with higher reliability.33PubMed Central. Challenges and prospects for advanced packaging

Only the parts that need top speed use the newest node, which can lower cost and improve yield, because a smaller chip is less likely to contain a defect. It also means the node label on a product may describe only part of the chip.