How Are Computer Chips Made? Light, Chemistry and Absurd Precision

The raw material is quartz, reduced in a furnace to metallurgical silicon and then purified chemically to a level described as eleven nines, meaning fewer than one foreign atom in a hundred billion, because a stray atom in the wrong place changes the electrical behaviour of the material. That ultra-pure silicon is melted and a seed crystal is dipped in and slowly withdrawn while rotating, drawing up a single continuous crystal, a process invented by Jan Czochralski in 1916. The resulting ingot, up to two metres long and three hundred millimetres across, is a single crystal throughout, with every atom aligned to the same lattice. It is sliced into wafers less than a millimetre thick, and each wafer is then polished to a flatness measured in nanometres, because the printing process that follows has a depth of focus too shallow to tolerate anything else.

The central operation, repeated dozens of times, is photolithography, which transfers a pattern onto the wafer using light and a light-sensitive chemical. The sequence for each layer runs:

• •A thin film is deposited across the wafer, of silicon dioxide, metal, or another material depending on what the layer does
• •A photoresist, a chemical that changes solubility on exposure to light, is spun onto the surface in a layer a fraction of a micrometre thick
• •Light is projected through a mask carrying the circuit pattern, shrunk by a lens system by a factor of four, exposing the resist in the pattern required
• •The exposed or unexposed resist is dissolved away in a developer, leaving a stencil
• •Etching removes the material not protected by the stencil, usually with a plasma of reactive gases
• •Ion implantation fires dopant atoms into exposed silicon to alter its conductivity, and the remaining resist is stripped
• •Repeat, aligning each new layer to the previous ones to within a few nanometres, for anywhere from sixty to a hundred layers

The limit on how small a feature can be printed depends on the wavelength of the light, and for decades the industry used ultraviolet light at 193 nanometres while printing features far smaller than that, using a series of increasingly ingenious tricks: immersing the lens in water to raise the refractive index, printing a pattern twice with a slight offset to halve the effective pitch, and reshaping the mask to compensate for the distortions diffraction introduces. Around the 7 nanometre generation these ran out, and the replacement is extreme ultraviolet at 13.5 nanometres, which took two decades and billions of dollars to develop. It is a remarkable machine: a droplet of molten tin is hit twice by a high-power laser, once to flatten it and once to vaporise it into a plasma that emits the required wavelength; that light is absorbed by air and by glass, so the whole system runs in vacuum and uses mirrors rather than lenses, each built from dozens of alternating layers and polished so precisely that if one were scaled to the size of Germany its largest bump would be under a millimetre. One Dutch company, ASML, makes them, each costs upward of a hundred and fifty million dollars, and they are the tightest bottleneck in the global supply of advanced chips.

A fabrication plant costs between ten and twenty billion dollars and is built around contamination control, because a single particle landing on a wafer destroys the chips beneath it. The air in the cleanest areas is filtered to fewer than ten particles above a tenth of a micrometre per cubic metre, staff wear full-body suits that protect the wafers from them rather than the reverse, and the buildings sit on massive isolated foundations because vibration at the nanometre scale matters. Wafers travel between hundreds of machines in sealed pods on overhead rails, and a single wafer takes two to three months to complete its journey through the plant. Yield, the proportion of chips on a wafer that work, is the number the whole economics turns on, since a new process may start below fifty percent and must be driven up over months of statistical detective work; defective chips are identified by testing every one on the wafer before it is cut, and many products are sold as lower-tier parts by disabling the sections that failed.

Gordon Moore's observation in 1965 that the number of components per chip was doubling roughly every year, later revised to every two, held for half a century and drove the industry's planning. It has slowed. Node names such as 5 nanometre or 3 nanometre no longer correspond to any physical dimension and have been marketing labels for over a decade, the actual gate lengths being considerably larger. Progress now comes less from shrinking features than from changing the structure, moving from flat transistors to fin-shaped ones around 2011 and to gate-all-around designs in current generations, and from stacking: chips are being built vertically, and complex products are assembled from several separately manufactured pieces bonded together, since a large monolithic die has a poor chance of being defect-free. The costs of each generation have risen enormously, and the number of companies able to build at the leading edge has fallen to three, which is why chip manufacturing has become a matter of national industrial policy in a way it was not twenty years ago.

Chips begin as silicon purified to fewer than one impurity atom in a hundred billion, grown as a single crystal and sliced into polished wafers. The pattern is printed by photolithography, projecting light through a mask onto a light-sensitive coating and then etching and doping, repeated for sixty to a hundred layers aligned to within nanometres. Printing the smallest features needs extreme ultraviolet light made by vaporising tin droplets with a laser, in machines built by one company, and progress now comes from transistor structure and stacking rather than from shrinking.