Starting with sand, ending with a chip
A modern processor contains billions of tiny switches called transistors, packed onto a piece of silicon often smaller than a postage stamp. Building it is among the most demanding manufacturing processes humans have developed. This article gives a simplified picture of the main stages. Real processes involve hundreds of steps and details that companies keep private, so treat this as an outline rather than a recipe.
Silicon and wafers
Silicon is a semiconductor, a material whose ability to conduct electricity can be controlled, which makes it suitable for switches. It is abundant, found in sand and rock, but chip-making needs it in extremely pure form. The silicon is melted and grown into a large single crystal, shaped like a thick cylinder. The cylinder is sliced into thin round discs called wafers, which are then polished until they are extremely smooth. A wafer is the canvas on which many identical chips are built at once, each chip being called a die.
Cleanrooms
Features on a chip are far smaller than a speck of dust, so a single particle can ruin a circuit. Fabrication plants, often called fabs, therefore work in cleanrooms, where the air is filtered continuously and staff wear protective suits. Temperature, humidity and vibration are controlled too. This is part of why a fab costs so much to build and run.
Building up layers
A chip is made from many layers stacked on the wafer. The process repeats a cycle of steps, with different materials and patterns each time.
Deposition. A thin film of material, such as an insulator or a metal, is added across the wafer’s surface. Several techniques exist, including growing films from gases.
Photolithography. This is the step that defines patterns. The wafer is coated with a light-sensitive material called photoresist. Light is shone through a pattern, called a mask or reticle, and focused down onto the wafer, usually shrinking the pattern. Where the light hits, the resist changes chemically and can be washed away, leaving a stencil. Producing the smallest features has required very short wavelengths of light, and the most advanced equipment today uses extreme ultraviolet light, which demands highly specialized and costly machines.
Etching. With the stencil in place, chemicals or plasma remove material from the exposed areas, carving the pattern into the layer beneath. The remaining resist is then stripped off.
Doping. To make transistors, regions of silicon are deliberately altered by adding small amounts of other elements, which changes how those regions conduct. One common method fires charged atoms into the silicon, a process called ion implantation.
Planarizing. Between steps the surface is often polished flat with a chemical and mechanical process, so the next layer can be patterned accurately.
Wiring it together
Once transistors exist, they have to be connected. Many layers of fine metal wiring, separated by insulating material, are built above the transistors. Tiny vertical connections link one layer to the next. In a complex processor there can be many such wiring layers, which together route signals and power across the chip. Copper is widely used for this wiring, though materials and methods vary.
What “nanometer” means
Chip generations are often labeled with a number in nanometers, such as a “3 nm process.” A nanometer is a billionth of a meter. These labels once corresponded fairly directly to the size of a feature in the transistor, but today they are largely marketing names for generations, and they do not measure a specific dimension in the same way across companies. Comparing two makers by the number alone can be misleading. Transistor shapes have also changed over the years, from flat designs to raised, fin-like structures and newer arrangements, as flat designs became harder to control at small sizes.
Testing the wafer
After processing, each die on the wafer is tested electrically using fine probes. Some dies fail because of defects. The share of working dies is called the yield. Yield matters hugely: a process with poor yield wastes expensive wafers, and larger dies are more likely to contain a defect, which is one reason very large chips cost disproportionately more. Some chips that partly work may still be sold as lower-tier versions, with faulty sections switched off.
Cutting and packaging
The wafer is then cut apart into individual dies. Each good die is mounted in a package: a protective housing with connections that link the tiny contact points on the die to the larger pins or pads that fit on a circuit board. Packaging is also a place for innovation. Some products combine several smaller dies in one package, sometimes called chiplets, which can improve yield and allow different parts to be made with different processes. Further testing follows, including checks at different temperatures and speeds.
A global, specialized supply chain
Few organizations do every stage. Some companies design chips but do not manufacture them, contracting the work to specialist foundries. Others make equipment, chemicals or ultra-pure materials, and the machines used in a fab often come from a small number of suppliers. Because advanced fabs are enormously expensive and take years to build, only a handful of firms worldwide can make the most advanced processors. That concentration has made chip supply a matter of attention for businesses and governments, particularly after shortages in recent years. Whether and how production should be spread more widely is a continuing policy debate.
Graphics processors, which are central to modern AI work, come out of these same fabs, and researchers also explore alternatives such as neuromorphic computing and photonic computing that would need their own manufacturing methods, though how far they will go is uncertain.
Why it matters for progress
The ability to etch ever smaller patterns, while keeping defects rare, is what has driven improvements in computing over decades. As shrinking becomes harder and more expensive, the pace of those improvements is changing, a story covered in the discussion of Moore law. The chips that train and run neural networks rely on the same manufacturing steps, which is why fab capacity affects the supply of AI hardware as well as phones and cars.