The Incredibly Complicated, Extremely Small World of Semiconductor Manufacturing

semiconductor microchip shortage

In 1965, a remarkably prescient young chemist working at Fairfield Semiconductor wrote an essay for “Electronics,” a trade magazine. His essay was quite technical, but it has been boiled down to a rule that has been often referenced over the years when people talk about the speed of technological advancement in computers. That chemist was Gordon Moore, the (eventual) co-founder of Intel, and the “rule” is known as Moore’s Law. In his essay, Moore predicted that the number of transistors in a dense integrated circuit (IC) would double about every two years. (Originally, it was actually one year, but he later revised his statement.) In layman’s terms, it means that computer speed doubles every two years, while the cost is halved. More computing power means not just more speed, but also more capacity, for a practically infinite number of applications.

When Moore talked about dense integrated circuits, he was talking specifically about semiconductors (colloquially also called a “chip”), the brain that powers pretty much all technology today. There are different types of chips that have other elements integrated into them, including central processing units (CPUs). Semiconductors with a CPU integrated into it are known as microprocessors, while chips with a CPU, as well as memory, a clock and an I/O unit are microcontrollers. The invention of the microprocessor was what made personal computers even possible.

As chips get smaller and more sophisticated, the manufacturing processes also become more complex and difficult. At the same time, consumer demand has skyrocketed.

There is fierce competition between manufacturers to build chips with smaller and better transistors. The first microprocessor, built by Intel in 1971, contained 2,300 transistors, at a node size of 10 microns each. Moore’s essay contemplated a single chip with 65,000 transistors. Today’s chips, no larger than your fingernail, have billions. The latest chips now have transistors under 10 nanometers in width, with the latest achievement being sub-5 nanometers (more specifically, that’s the width of the logic gate in a transistor). This is so small, the human mind can’t really comprehend it. A person’s fingernail grows at a rate of 1 nanometer per second. Either a sheet of paper or a strand of human hair is 100,000 nanometers thick. A COVID-19 virus particle (not a technical term) is approximately 100 nanometers.

The manufacturing technique used to actually make chips with this many transistors is called photolithography. Lithography has actually been around for thousands of years. It’s a printmaking process in which a design is etched onto something using a chemical process that leaves ink or stain only in the places the designer wants it. It’s kind of like a stencil, except instead of a physical barrier to block the ink, there’s a chemical barrier that rejects the ink instead.

Photolithography is a similar process, only light is used instead of ink. In photolithography, light is used to etch intricate patterns onto thin layers of substrate (silicon wafers). The light is focused onto a mask, which is essentially a stencil of the final pattern, then through a series of optical lenses, the image is shrunk down. Many layers of silicon, with different designs, are carefully aligned into a 3D pattern, building the integrated circuit that incorporates billions of transistors.

When creating the chip, the light source determines the intricacy of the patterns on the wafers; the key to getting more transistors onto smaller chips is the light’s wavelength. A shorter wavelength means more transistors. The development of more advanced light sources for photolithography, which is capable of mass manufacturing, is one of the key innovations needed to keep Moore’s Law true.

Until a few years ago, the most sophisticated photolithography available used deep-ultraviolet (DUV) light, which has a wavelength of 193 nanometers. A problem with this process is that as the lightwaves pass through the optical lenses, which are made of glass, the lenses actually absorb the light — so by the time you’ve focused the light enough to get to 100 nm and below size, there’s no light left to actually etch the pattern onto the silicon wafer.

A few next-generation lithography techniques were possibilities, and many of them have been in development since the late ‘80s or early ‘90s. Most were abandoned after years of research and billions in investment, leaving only one option left: extreme ultraviolet lithography (EUVL). The only company stuck with the development of EUVL was ASML, a Dutch company spun off of Philips in 1984.

EUVL works in a similar manner to deep-ultraviolet, except that mirrors are used instead of lenses, to create a reflective mask. It also has a much shorter wavelength of 13.5 nm, significantly shorter than previous techniques. EUVL has had many barriers to overcome, not the least of which was finding a light source powerful enough for the process to work (chipmakers wanted 250 watts of power). The lithography also has to take place in a vacuum; the light wavelengths are so short that even air absorbs them. The mirrors also have to be perfectly flat, with no defects, or the optics can be affected, altering the final pattern.

ASML showed a working prototype of its EUVL machine in 2001, but it wasn’t until 2018 that ASML finally shipped the first one for actual commercial production. It went to Samsung. Now there are over 100 EUVL machines being used to produce next-generation semiconductors. One of ASML’s machines costs over $150 million, and it’s huge, taking up 40 freight containers when shipping. There’s a waiting list.

For the past 20 years or so, there have been many articles and speeches proclaiming the death of Moore’s Law. Moore’s Law is bumping up against the laws of physics. Even Gordon Moore himself is skeptical how much longer it can continue. A lesser-known companion to Moore’s Law, Rock’s Law, states that the cost to build a semiconductor chip fabrication plant doubles every four years. If physics doesn’t kill Moore’s Law, then perhaps economic viability will.

But for now, ASML is already working on the photolithography equipment that will come after EUVL, called “High-NA EUV.” NA stands for numerical aperture. This machine will allow for 3nm chips. The first of these isn’t expected until 2025 at least (and will cost at least $300 million), but it’s a promising technology.

Other techniques have actually pushed DUV beyond original expectations for chip miniaturization, so it will remain in use for the foreseeable future. Companies are researching other ways besides just new photolithography techniques to increase transistor counts on chips. IBM has been working on a new integrated circuit design called nanosheet, which the company says can fit 50 billion transistors onto a 300mm chip.

So in the meantime, and in particular for those of us working in technology, we can only hope that the rumors of the death of Moore’s Law have just been greatly exaggerated.