How Are Chips Made? The Semiconductor Manufacturing Process

In everyday life

Phones and computers contain several small black parts that compute or store data. Inside each one is a "chip": a tiny piece of silicon with a circuit built into it. The chip that does the computing in a phone holds billions of tiny switches called transistors. "What Is a Transistor?" explains them.

Something this small cannot be put together one piece at a time. Instead, many chips are made together on one plate and cut apart at the end.

The big picture: three stages

Making a chip can be divided into three broad stages. First, a round plate called a wafer is made from silicon. Next, transistors and wiring are built on the wafer. This stage is called the front end.

Finally, the wafer is cut into individual chips, which are finished as parts and tested. This stage is called the back end. The table in "Front-end vs. back-end process" also shows how the two differ.

From rock to wafer

Silicon is taken from silicon dioxide (silicon bonded with oxygen), which is plentiful in sand and rock. But beach sand is not used as it is. High-purity quartzite (a rock made of quartz) is heated with carbon in an electric furnace to remove the oxygen. The silicon made this way is still only about 98% pure.

Next, it is turned into a liquid compound to remove impurities, then turned back into solid silicon (the Siemens process). The result is polycrystalline silicon, made of many small crystal grains. Its purity rises to about 99.999999999%, or "eleven nines." "What Is a Semiconductor?" explains why such purity is needed.

Chips need a single crystal, in which the atoms are lined up in one pattern throughout. So the polycrystalline silicon is put in a heat-resistant quartz container and melted at a little over 1,400 °C. A small piece of single crystal (a seed) is dipped in and slowly pulled up while rotating. The crystal grows thicker and longer, keeping the same pattern as the seed. This method is called the Czochralski (CZ) method.

Because it rotates as it is pulled up, the crystal comes out shaped like a cylinder. This block is called an ingot. Wafers are round because they are slices of this cylinder.

The ingot is cut into round plates less than 1 mm thick, by a machine that cuts with thin wire. The surface is then polished as flat as a mirror, and the wafer is done. The largest wafers used in mass production are 300 mm in diameter.

The front end: print a pattern, then etch

In the front end, hundreds of identical chips are made on one wafer, laid out in a grid. Most steps work on a whole wafer at once. Chips are not processed one at a time, so many chips can be made with the same quality.

The core is a series of steps that carve a circuit pattern into a film. They go in order from ① to ⑥ in the figure.

1. Add a film (deposition): cover the whole wafer surface with a very thin film. It may be a film that does not conduct electricity, or a metal film that will become wiring.
2. Coat with resist: spread a thin layer of a chemical that changes when light hits it on top of the film.
3. Expose to light (exposure): shine light through a master plate that carries the circuit pattern. The resist changes in the shape of the pattern.
4. Develop: use a liquid to dissolve the resist where light hit it. (In some types, the parts that light did not hit dissolve instead.)
5. Etch: remove the film that is not covered by resist, using gases or liquids.
6. Strip the resist: remove the resist, which has done its job. The circuit pattern is left in the film.

The chemical coated on in ② is called photoresist. In the fab, it is simply called "resist." There are two kinds: one where the parts hit by light dissolve, and one where they stay.

The master plate used in ③ is called a mask. It is a glass plate with the circuit pattern drawn in a film that blocks light. The exposure machine shrinks this pattern with lenses or mirrors and projects it onto the wafer. It prints only part of the wafer at a time. So it shifts the wafer bit by bit, printing the same pattern across the whole surface.

Steps ② to ④, which use light to copy a pattern into resist, are together called lithography. The word was first the name of a printing method that used stone plates. How fine a pattern lithography can print largely decides how fine a chip can be made.

Other front-end steps

Besides printing and etching patterns, the front end has several other important steps. Here are three typical ones.

• Adding impurities: elements that set a semiconductor's properties (impurities) are sped up by electric force and shot into the wafer. This step is called ion implantation. They cannot enter areas covered by resist, so conductivity changes only where it is meant to.
• Polishing flat: stacking layers leaves the surface bumpy. If bumps remain, parts of the next exposure will be out of focus. So the surface is polished flat with a chemical liquid and a polishing pad. This step is called CMP.
• Cleaning: even tiny particles or traces of metal can make a circuit fail. So the wafer is cleaned before and after many steps. Cleaning is one of the most frequent steps in the front end.

Along the way, film thickness and pattern sizes are also measured, and wafers are checked for scratches and particles. The earlier a problem is found, the fewer wafers are wasted.

Why repeat it dozens of times?

A chip is not made from a single pattern. It is a 3D structure, built up from many layers, from the bottom up. Each layer has a different shape. So each layer uses its own mask, and the steps above are repeated.

First, transistors are built on the silicon surface. The transistors alone need many masks: for where impurities go, for the electrodes that apply voltage, for the films that separate them, and more. Then the wiring that connects the transistors is stacked in many layers, with films that do not conduct in between. Wiring stacked in many layers like this is called multilevel interconnect.

Billions of transistors cannot be connected with a single flat layer of wiring. The lower layers use thin wires to link nearby transistors. The higher the layer, the thicker the wires, which link distant parts of the chip and carry power. The layers are connected to each other by thin vertical wires, like pillars.

Each new layer's pattern must line up exactly with the pattern below it. The exposure machine reads marks on the wafer and lines it up before shining the light. The number of repeats depends on the type of chip. ASML, a maker of exposure machines, says the pattern-printing step is repeated more than 100 times.

Cleanrooms: why dust is the enemy

The front end takes place in rooms where tiny particles in the air are cut down to an extreme degree. Such a room is called a cleanroom. Air flows down from the ceiling through filters that catch even very fine particles. It constantly pushes particles down toward the floor.

Dust is a problem because chip patterns are extremely fine. A nanometer (nm) is one millionth of a millimeter. An IEEE Spectrum article (2020) says the fine wires in the most advanced chips of the time were spaced about 36 nm apart. More than 2,000 of these wires would fit across the width of one human hair (about 0.09 mm thick).

At this scale, a single 100 nm particle can land across several neighboring wires. The wires may become connected or broken, and that chip becomes defective. A particle too small to see is like a large boulder to a chip.

How clean the air is gets graded in classes set by an international standard (ISO 14644-1). City air outdoors is about class 9, the loosest class. In class 1, the cleanest, there may be no more than 10 particles of 100 nm or larger per cubic meter (m³) of air. Fabs control the areas that wafers pass through most strictly, and carry wafers in sealed containers.

Workers wear special suits that cover the whole body. Human bodies and clothes constantly shed particles, such as flakes of skin and fibers. These suits are less about protecting people and more about protecting wafers from the particles people shed.

The back end: cut, package and test

A wafer that has finished the front end is covered with completed chips. But not all of them are good. So first, while the chips are still on the wafer, fine needles touch each chip's electrodes to check that it works correctly. This test is called a wafer test, and the positions of bad chips are recorded.

Next, the back of the wafer is ground thinner. Then the wafer is cut into individual chips, for example with a spinning blade. This step is called dicing. Each chip is then fixed to a base (a substrate or a metal frame) that will connect it to the outside.

There are two main ways to connect the chip's electrodes to the base. One is wire bonding, which links them one by one with gold or copper wires thinner than a hair. The other, called flip chip, turns the chip upside down. Tiny metal bumps on its electrodes then connect it directly. "Wire bonding vs. flip chip" compares the two.

The connected chip is sealed in resin or a similar material, and terminals for outside connections are added. Now it has the form of a part. The whole thing, including this enclosure, is called a package. The package also protects the chip from moisture and shock, and helps release heat.

Finally, the finished part is tested once more. This is called the final test. It checks that packaging did not damage the chip, and that the chip performs at the specified temperatures and voltages. Only parts that pass are shipped. "Wafer test vs. final test" sums up how the two tests differ.

How long does it take?

A 2021 explainer by the Semiconductor Industry Association (SIA), a US industry group, gives these times. The front end takes about 12 weeks on average. Advanced chips can take 14 to 20 weeks. Assembly and testing in the back end take about 6 more weeks.

The front end takes a long time because similar steps are repeated for every layer, as we have seen. One machine is not enough: wafers travel back and forth between many machines. So even if orders suddenly rise, shipments cannot rise right away.

Yield: not every chip is good

However carefully a fab is run, a certain share of the chips on a wafer will be defective. The share of good chips among those made is called the yield. It strongly affects the price of each chip. If small defects are scattered at random, a bigger chip is more likely to hit one, so the yield goes down.

For example, a 300 mm wafer gives about 600 chips that are 10 mm square. With 0.1 defects per cm², about 90% of them are good. At 20 mm square, you get about 150 chips, and only about two-thirds are good (rough estimates from this site's calculator). Try different chip sizes with the die per wafer and yield calculator.

Who makes chips?

A single company rarely handles this whole flow. The work is divided up, and each stage is taken on by companies that are good at it. The table below shows, from left to right, what each kind of company does, what the industry calls it, and examples.

You don't need to remember every name. The one you will see most often in the news is the foundry, which manufactures chips for others. "IDM vs. fabless vs. foundry" compares them.

Common misconceptions

By the numbers

Put side by side, these numbers show how hard the front end is. Patterns spaced about 1/2,000 of a hair's width are stacked dozens of times, each layer lined up with the last. All the while, not a single particle may land on any of the hundreds of chips.

Learn more

• Compare in tables: CZ vs. FZ silicon (ways to grow single crystals), CVD vs. PVD (ways to add films), Wet vs. dry etching, Positive vs. negative photoresist
• Try the numbers: the lithography resolution calculator shows how the wavelength of light changes how fine a pattern can be printed.
• See it in figures: the exposure and etching steps, CZ crystal pulling, a flip-chip cross-section and more are in the list of figures.