From sand to wafer
Every chip starts as a slice of one perfect crystal: a single lattice of silicon atoms as long as a person is tall, grown slowly out of a pot of liquid silicon.
It starts as quartz rock.
The silicon in a chip starts as quartz or quartzite. Mixed with coal, coke, charcoal or wood chips and heated to about 1,500 to 2,000 °C in an electric arc furnace, the quartz gives up its oxygen, and liquid silicon is drained from the bottom of the furnace. It comes out 98 to 99% pure: metallurgical-grade silicon, which in the United States goes mostly into aluminium alloys and silicones.
Most of it is made in China.
USGS estimates that the world made 4.6 million tonnes of silicon metal in 2025, and China 4.0 million of them, 87%. The next three are Brazil, Norway and France. In Europe, Elkem mines its quartz in Norway and Spain, and Wacker makes its own silicon metal at Holla in Norway. Pick a country.
Counting the nines.
Purity is counted in nines: 99.999999999% is 11 nines, which the industry writes 11N. Silicon for chips is 11 nines or purer: Shin-Etsu and Tokuyama state 11 nines for theirs, and Wacker 12 nines. Solar silicon is 6 to 10 nines. At 11 nines, one atom in a hundred billion is something other than silicon.
This cube holds 27,000 atoms. Slide the purity up and watch the red ones, the foreign atoms, disappear long before the chip industry is satisfied. Then read the last box: even at 11 nines, one wafer still holds about 27 trillion of them.
Growing rods out of a vapour.
To reach the nines, silicon metal is turned into a liquid, trichlorosilane, distilled, and turned back into silicon. In the Siemens process, thin U-shaped silicon rods in a steel bell jar are heated electrically, up to 1,150 °C, and trichlorosilane heated to over 1,000 °C passes over them. It leaves pure silicon behind, layer on layer, until the rods are 15 to 20 cm thick.
The process was developed at Siemens in Pretzfeld, Germany, between 1946 and 1956, and Wacker began making polysilicon on an industrial scale in 1959. Most polysilicon now goes into solar cells: for 2025, TECHCET forecast demand for the semiconductor grade at 33,500 tonnes, 2.4% of the total.
Pulling a crystal from a pot.
The crucible is where the sand is: it is made from high-purity quartz sand. Polysilicon chunks go into it with boron or phosphorus, which set how the silicon will conduct, and melt at around 1,420 °C under argon. A seed crystal touches the surface and is pulled up while turning, and the whole ingot grows as one crystal in the seed's orientation.
Siltronic calls its crystals man-sized: around 300 mm across, several hundred kilograms, five to seven days in the puller. Jan Czochralski, a Polish chemist who worked at AEG in Berlin from 1907, described the method in a paper he handed in on 19 August 1916.
A seed: a pencil-thin crystal with the exact lattice orientation wanted, dipped into the melt.
Neck: pulled fast and thin first, so that defects from the thermal shock grow out and the crystal starts perfect.
Shoulder: the pull slows and the crystal widens until it reaches its full diameter.
Body: the long, straight part that becomes wafers. It grows slowly, the crystal turning as it rises.
Finished: the crystal is lifted clear. Everything in it was once a pile of polysilicon chunks.
The nines belong to the polysilicon. As the crucible's quartz slowly dissolves, oxygen enters the crystal: 11 to 17 parts per million atoms in GlobalWafers' measurements, far above the nines. Wafer makers specify it rather than remove it: GlobalWafers writes that it strengthens the wafer and traps metal contamination.
A saw made of wire.
The crystal is cut into blocks and ground to an exact diameter, and a notch is ground along its side to mark the crystal's orientation. Every wafer cut from it carries that notch.
Then a wire saw slices the whole block at once. The block is lowered into a web of fast-moving wire, one single wire fed from spool to spool, and an abrasive slurry poured on the web does the cutting. SUMCO cuts slices of around 1 mm.
Flatter than anything you own.
The slices are rounded at the edge, lapped between two plates turning in opposite directions to take about ten micrometres off each face, and etched to remove the damage that is left. A 300 mm wafer is then polished on both faces at once, held, as in lapping, between an upper and a lower polishing plate, with a slurry between them.
Flatness matters because a lithography scanner's depth of focus has shrunk as its wavelength got shorter and its lens aperture larger. The industry's roadmap asked for 300 mm wafers flat to about 14 nm by 2020 across each 26 by 8 mm exposure site. Siltronic's picture of 20 nm: a flat sheet on the water of Lake Chiemsee.
Ready for a thousand steps.
The finished wafer is 300 mm across and 775 µm thick. The industry calls it a 12-inch wafer, about the width of a 12-inch record, and it is about as thick as a bank card. It weighs about 128 g, by our arithmetic. It travels to the fab in a shipping box such as ePAK's, which holds 25 wafers 10 mm apart.
Wafer makers on this record: GlobalWafers, Shin-Etsu, Siltronic, SK Siltron and SUMCO. Polysilicon makers: Hemlock, Tokuyama and Wacker.
- Anatomy of a fabthe building the wafer goes into02
- Writing with lighthow each layer is printed03
- What a chip costshow many chips one wafer is worth10