Silicon Planet: My Life in Computer Chips

By Pat Hays

Computer chips, like sand and water, now cover the earth. Each year, more than 60 billion computer chips are shipped inside smartphones, data centers, cars, airplanes, refrigerators, and more. How did we get to this point? And what even is a computer chip?

For nearly forty years, author Pat Hays developed state-of-art computer chips, as a designer and entrepreneur. Silicon Planet is his personal story, full of the ideas, people, and drama at the heart of this technology. In it, the reader will learn how computer chips are designed and fabricated while witnessing breakthroughs and battles that have shaped our world.

Silicon Planet concludes with an analysis of the Boeing 737 MAX disasters—events that exemplify the risks to safety and security that computer chips have brought alongside their benefits. Silicon Planet will engage general readers as well as engineering professionals.

• Language: English
• Publisher: Pat Hays
• Release date: Apr 4, 2023
• ISBN: 9798987644119

Pat Hays

Pat Hays enjoyed a long career as a leading developer of computer chips. Hundreds of millions of computer chips that he designed or managed have been shipped. Pat’s chips were often among the first to use programmable processing for telecommunications, consumer products, security, and other emerging applications. He is a past recipient of the Best Paper Award at the International Solid-State Circuits Conference, the foremost global forum for chip advances. Pat received his PhD in physics from MIT.

Book preview

Part 1

Into the Cathedral

1.1

Silicon Everywhere

Bicycling the four miles west from my home in Sarasota, through town, then across the Ringling Bridge, hill training for a Florida cyclist, I can avoid all traffic congestion and I can park wherever I like, but I always park in the bike rack where the road ends at the north end of Lido Key. I hop off my bike, a beater that I leave outside year-round, and cable-lock it to the rack.

I walk several hundreds of feet to the beach, as it widens into mounds of wild grasses and sea grapes. Americans arrive at the beach wheeling carts that overflow with chairs, umbrellas, food, and toys, as if they’re emigrating. A Brazilian wears sandals, a towel over their shoulder, and a cell phone in the waistband of their bathing suit. I want to be like the Brazilians. I carry a small backpack with a towel, a book, water, and, common to both Americans and Brazilians, a cell phone.

Today, black skimmers are nesting on the beach and I need to weave among them on my way to the shore. In my recent retirement to Sarasota, most of my Instagram posts are of shorebirds, egrets, ibises, and diving pelicans. Yet sometimes, when I feel the soft white sand between my toes, I think of silicon.

Silicon is number 14 in the periodic chart of the elements, the chart that the Russian chemist Dmitri Mendeleev formulated in 1869 to organize the known elements according to their properties. We now know that the silicon atom has 14 protons, orbited by 14 electrons. Like the other elements in the same column of the periodic chart, carbon, germanium, tin, and lead, four of the silicon electrons, the valence electrons, are in an unfilled outer orbital band. An element’s valence electrons can be shared with other atoms to form filled orbital bands, creating chemical compounds with preferred, lower energy states. As a result, pure silicon is rarely seen in nature because it readily bonds with oxygen to form silicon dioxide, a three-dimensional tetrahedral structure.

Silicon dioxide is the primary constituent of sand. It covers the earth’s surface, making silicon the second most abundant element on earth by weight, after oxygen. In English, silicon dioxide is called silica; in French, silice; in Russian, kremnezem; in Icelandic, kísil. From ancient times, silica has been used to form concrete; the Roman Pantheon dome (113–125 AD) was built of concrete. During the Han dynasty (25–220 AD), silica was refined to porcelain at temperatures between 2200 and 2600 Fahrenheit. And a small amount of silicon, highly purified from silica, is used to fabricate semiconductors.

Semi is borrowed from Latin, meaning half. From that you may infer that a semiconductor is a half conductor of electricity. In fact, at room temperature, pure silicon is a conductor, a weak conductor, roughly halfway between glass, an insulator, and copper, a good conductor. With impurities, silicon, depending on the type and amount of the impurity, becomes a better conductor.

The invention of the first transistor in 1948 is generally credited to the physicist William Shockley, who, with John Bardeen and Walter Brattain, led a world-class team at AT&T Bell Laboratories. They sandwiched two slices of germanium, implanted with impurities, on top of each other. They demonstrated that a small current from a gold contact on the tops of the slices could be used to control a larger current between the two germanium slices.¹

The three men shared the 1956 Nobel Prize in Physics for their researches on semiconductors and their discovery of the transistor effect.² Shockley himself inspired the old saw that Silicon Valley’s greatest resource is its terrible CEOs. In 1956, he left Bell Labs to found Shockley Semiconductor in Palo Alto. Within a year, his mismanagement drove eight of his top staff to quit and establish Fairchild Semiconductor, a new division of Fairchild Camera and Instrument. By 1968, friction with upper management caused Gordon Moore and Robert Noyce to leave Fairchild to cofound Intel.

Shockley’s personal story became much worse in the mid-1970s. In a series of articles and lectures, he claimed to prove the genetic inferiority of Black people. After that, despite his Nobel Prize, Bell Labs tried to overlook Shockley’s name whenever possible. Fortunately, Bell Labs had a deep bench.

Shockley’s device, a point-contact transistor, was fragile and costly. During the 1950s two Bell Labs engineers, Mohamed Atalla and Dawon Kahng, developed a new type of transistor called the metal-oxide-semiconductor field-effect transistor (MOSFET).³ Atalla and Kahng implanted impurities (ions) in two disjoint regions in a single silicon substrate. Next came the critical invention, silicon passivation, that exploited the affinity of silicon for oxygen in order to grow an insulating layer of silicon dioxide on top of the semiconducting substrate. This layer protected the silicon and stabilized its electrical properties. When voltage was applied above the gap between the two regions, electrons flowed between them, a phenomenon called tunneling. In the simplest case, the control voltage could be used to turn the current between the two implant regions on or off, and the MOSFET functioned like a tiny on/off switch. Compared to Shockley’s transistor, Atalla’s transistor was reliable and it could be mass-produced at low cost.

Figure 1-1. Cross Section of a MOSFET Transistor

In the late 1950s, Jack Kilby at Texas Instruments, followed soon after by Robert Noyce at Fairchild, demonstrated techniques to integrate multiple, interconnected transistors onto a small chip of silicon. These devices were called integrated circuits (ICs). As more transistors were integrated, the term large-scale integration (LSI) came into vogue, and, by the 1970s, very large-scale integration (VLSI). Later, the term ultra-large-scale integration (ULSI) was occasionally used, but, generally, the naming stopped with VLSI, now a synonym for almost any chip.

In 1965, the Intel cofounder Gordon Moore observed that the transistor density of semiconductors was doubling every year and forecast that trend to continue. Ten years later, he updated his now-famous Moore’s law to forecast that the doubling rate would flatten from annual to biannual after 1980. Until the last several years, the updated Moore’s law has held remarkably steady.

To an ex-physicist like me, it’s strange to call this trend a law. Is it a law of nature, like gravity? Moore’s law started as an empirical observation, subject to significant error. The predicted growth became self-fulfilling: it set the goals that semiconductor companies had to meet to stay competitive. Starting in 1993, these goals were formalized and documented by the Semiconductor Industry Association with its International Technology Roadmap for Semiconductors (ITRS). By that time, the industry had become so complex and disaggregated that numerous independent developments wouldn’t result in consistent quantum steps forward without a common specification, at least a straw man. Fulfilling the ITRS specifications required a series of inventions, which the industry delivered on schedule year after year.⁴

As the surface of the earth is covered with sand, it’s now covered with semiconductors. In 2018, for the first time, over one trillion chips were fabricated and shipped.⁵ In 2021, Apple’s M1 Max, the central processing unit (CPU) for its recent desktop computers, integrated 57 billion transistors on a single chip.⁶ Total unit-shipment counts are skewed toward low-end semiconductors with fewer transistors. If we estimate that the average chip had a mere 10 million transistors then, in 2019, 10 quintillion (10,000,000,000,000,000,000) transistors shipped; that is, over one billion transistors per year for every man, woman, and child on earth. That number may sound implausible. I thought so too until my first video call that year, from Omar, my camel driver in Hassi Labiad, a dusty village of casbahs in the northern Sahara, five klicks from the road’s end in Merzouga, in southern Morocco near the Algerian border.

Of the trillion chips shipped in 2019, roughly 60 billion are classified by market analysts as computer chips, chips that can perform different tasks, depending on their software. The computational units, memories, and control systems at the heart of the computer chips are called programmable processors, or processors for short. Although only 6% of the total by unit count, computer chips command a highly disproportionate share of the $555 billion annual semiconductor revenue.⁷ For good reason: they are the queen bees among semiconductors. Memory chips store programs and data for processing by computer chips; sensor chips, like the accelerometer in the cell phone, collect data for the cell phone’s central computer chip, the application processor, to process and display; analog chips, another major category, typically convert real-world signals like sound waves into digital form for processing by . . . well, you get the idea.

Where did 2019’s 60 billion computer chips end up? Where are all those computers? In 2019, 260 million personal computers were shipped.⁸ In recent years, a lot of processing power has migrated from the desktop and laptop to vast data centers owned by Amazon, Microsoft, Google, and Facebook. By my estimation, data centers account for another hundred million computer chips each year.

Let’s not forget the big computers, the supercomputers. Though few in number, a supercomputer might incorporate over 100,000 interconnected nodes, each with a computer chip. Add a couple hundred million more computer chips to our total. Now let’s add smartphones; their computation power, for most tasks, is a notch below laptop computer chips only because the smartphone design is more power-constrained. About 1.4 billion smartphones shipped in 2019.⁹ At this point, we’ve still only accounted for a few billion of the 60 billion computer chips. Where are the other tens of billions of computer chips?

Computer chips are buried inside virtually every product: watches, cars, airplanes, refrigerators, printers. Computer chips are nothing less than the means for the product developer to implement functions that might in the past have been implemented with specialized hardware designs, or not implemented at all.

Computer chips that are programmed only by the manufacturer and not by the customer are embedded in contrast to the open computer chips in laptops and smartphones that are programmable by their customers. You probably didn’t write the software programs for your laptop or your smartphone. That fact is incidental; it’s feasible for both laptops and smartphones to be programmed by their customers, whether directly or indirectly by a third-party application developer.

Open computer chips are the tip of the unit-shipment iceberg; the embedded computer chips are below the sea, out of sight but accounting for far more units. The embedded computer chip markets are so numerous, it’s difficult to think of any product of even modest complexity that does not embed one or more computer chips. Perhaps, a Rolex mechanical watch? Correct, but insignificant.¹⁰ The Rolex Submariner, starting at $10,995 on Amazon, is more jewelry than watch. I’m satisfied to get the time from the computer chips in my cell phone⁠—and it’s more accurate than the best mechanical watch because it uses the Global Positioning System (GPS) to synchronize to satellites. If you don’t count my GoPro, there are no computer chips in my (nonelectric) bicycle. However, the manufacturing of my bicycle, as with the Submariner, required many computer chips.

In early 2021, the world learned that cars can’t run without semiconductors.¹¹ Today’s automobile carries over a hundred computer chips, ranging from chips that flash the taillights⁠—it’s cheaper to program a computer chip than to design a specialized taillight flasher⁠—to car navigation chips that can infer the outline of a winding road from its video image at 120 kph on Route 443, the Ma’ale Beit Horon, from Jerusalem to Tel Aviv.¹² The auto industry, anticipating a prolonged slowdown as a result of the Covid-19 pandemic, reduced its chip orders. When they realized their estimates were mistaken, it was too late. The production slots in the semiconductor fabrication lines had been reallocated.

The tide is retreating to reveal a sandbar in the water just off shore. Beachcombers are looking for shells. With every trip to the beach, I bring home a shell. Collectors are out before dawn; they look for perfect shells. I’m happy to find the ones they leave behind. I like the little holes that show where other mollusks have drilled in. A shell is formed in four years or so, and may support life for 10 years, then wash ashore the next day or after a thousand years.

When I retired in 2018 after 38 years of computer chip design, one of the transistors on my last design occupied one-millionth of the area of one of the transistors on my first design. A computer chip takes two to four years to design. It may find 10 years of useful application. No computer chip is perfect. By the time the design is finished, even before fabrication, its limitations would be painfully clear to me. All you can do is go on to the next one. If the computer chip is perfect, that’s the end of the game, my game. As the author Eve Babitz said, I have never liked perfect things, they give me the creeps.¹³

In my mind’s eye, past the shell collectors, beyond the grasses and mangroves and black skimmers, I see someone collecting sand, sand that will be processed into silicon, and from there into semiconductors. Excuse me for a minute, my cell phone is ringing.

---

1. November 17–December 23, 1947: Invention of the First Transistor, APS News 9, no. 10 (November 2000).

2. William B. Shockley, Nobel Prize, www.nobelprize.org/prizes/physics/1956/summary/ . In 1972, Bardeen became one of only four individuals ever to receive a second Nobel Prize. The second time, it was for his work on superconductivity.

3. Martin (John) M. Atalla, Dawon Kahng, National Inventors Hall of Fame, www.invent.org/ .

4. The ITRS abandoned its semiconductor roadmap following its last publication in 2017. New three-dimensional transistors starting with 14 nm line widths made identification of a common process node, according to its lithography, less and less meaningful.

5. Semiconductor Units Forecast to Exceed 1 Trillion Devices Again in 2021, IC Insights , January 24, 2019, www.icinsights.com/news/bulletins/Semiconductor-Units-Forecast-To-Exceed-1-Trillion-Devices-Again-In-2021 .

6. Andrei Frumusanu, Apple’s M1 Pro, M1 Max SoC’s Investigated, AnandTech, October 25, 2021, www.anandtech.com/show/17024/apple-m1-max-performance-review .

7. Global Semiconductor Sales, Units Shipped Reach All-Time Highs in 2021 as Industry Ramps Up Production Amid Shortage, Semiconductor Industry Association press release, February 14, 2022.

8. Gartner Says Worldwide PC Shipments Grew 2.3% in 4Q19 and 0.6% for the Year, Gartner press release, January 13, 2020.

9. S. O’Dea, Global Smartphone Shipments Forecast from 2010 to 2022, Statista, July 27, 2022, www.statista.com/statistics/263441/global-smartphone-shipments-forecast/ .

10. Neil Mawston, Apple Watch Outsells Entire Swiss Watch Industry in 2019, Strategy Analytics, February 5, 2020, www.strategyanalytics.com/strategy-analytics/blogs/wearables/2020/02/05/apple-watch-outsells-entire-swiss-watch-industry-in-2019 . According to this article, in 2019, Apple shipped 30.7 million watches versus 21.1 million for the Swiss watch industry.

11. Jack Ewing and Don Clark, Lack of Tiny Parts Disrupts Auto Factories Worldwide, New York Times , January 13, 2021.

12. Mobileye, based in Jerusalem, is a leader in car-navigation chips. Its CEO told me that he’d driven from Jerusalem to Tel Aviv without hands. Intel acquired Mobileye in 2017 for $15.3 billion.

13. Eve Babitz, I Used to Be Charming: The Rest of Eve Babitz (New York Review Books Classics, 2019).

1.2

On the Boulevard

Western Electric was founded in 1869, not long after the Civil War. After losing the fight for the patent rights to Alexander Graham Bell’s invention, it was acquired in 1881 and became AT&T’s manufacturing arm. The Western Electric building in Allentown was a block long, an 800,000-square-foot concrete hulk alongside the Union Boulevard sidewalk. It was painted yellow, the faded yellow of the old Sears buildings. To those in the know, the building was simply called Union Boulevard.

The Communications Act of 1934 granted AT&T a legal monopoly in exchange for end-to-end phone service for every US city, town, and farm. At the time of my visit in 1980, AT&T employed one million people. It was headquartered in a historic building at 195 Broadway in lower Manhattan. A 24-foot winged statue covered in 14-karat gold leaf, the Spirit of Communications, topped the building.

Bell Laboratories was the research and development arm of AT&T. AT&T’s monopoly gave its 15,000 engineers and scientists tremendous power. When they developed a product, it was going to roll out; marketing and sales didn’t get in the way. Engineers talked about deployment. They planned gigantic telecom systems 60 years into the future.

The three of us, my interviewer, the security officer, and I, walked a long, featureless hall parallel to the street until we reached a pair of glass doors. One door blocked our passage and led to a small chamber, with its second door on our far side. When the near door opened, my host and I, one at a time, entered the interlock. After a remote voice questioned us and inspected our images, we passed through the next door, met the next security officer, and continued our walk.

Union Boulevard housed one of the world’s most advanced semiconductor fabrication plants. The fab was capable of manufacturing semiconductors with circuitry as small as a single micron, about 1/100 the diameter of a human hair. It had been responsible for the world’s first commercial transistor production.¹⁴ I was ready to be awed.

At the end of the hall, we entered a locker room. We pulled on translucent plastic bunny suits and booties over our shoes, then face masks. This uniform was not intended to protect us, it was intended to protect the transistors from us. The Union Boulevard fab was a class 100 cleanroom.¹⁵ That meant that fewer than 100 0.5-micron-sized particles per cubic foot were tolerated. All equipment was sterilized, and downward airflow directed contaminants to the floor.

The locker room door opened into a room larger than a football field, filled with row after row of metal boxes, each the size of a car or small room. Engineers in bunny suits like ours peered intently into glowing windows and slots in the boxes, and a few carried trays of silicon wafers from box to box.¹⁶

A misset dial on one of the machines would destroy the wafers. The loss could kill a project, which in turn could kill a small company. Take a look around, my host said. No one here has ever made a mistake. After a mistake, an employee was reassigned to other work. I’d worked on my PhD thesis, a long calculation, with pencil and eraser on loose-leaf paper. As I filled pages and accumulated them into a three-ring binder, I often had to backtrack, tearing out 10 pages at a time. I made many mistakes.

My guide not only let me gawk but also explained what I was seeing.

Chip manufacturing starts with cylindrical crystals of pure silicon. Each cylinder is sliced across its vertical axis into circular wafers about 0.7 millimeters thick. The homogeneity of the crystal must be extremely high, which has constrained the diameter of the wafers. Today’s maximum diameter is 450 millimeters; however, 300 millimeters or smaller is often more economical. Wafer production is usually subcontracted to specialist companies and standardized across the industry so that the wafers are compatible with downstream equipment from different companies. The wafers arrive at the fab uniformly coated with silicon dioxide.

The fabrication process divides the wafer’s surface into identical rectangular sites. Within each site, the identical semiconductor design consisting of many transistors⁠—today, millions or even billions of transistors⁠—is simultaneously fabricated by depositing layers of different materials onto the wafer.

Early chips had only three electrically active layers, stacked on top of the wafer and sandwiched between layers of inert silicon dioxide. However, many steps were required to pattern these layers onto the wafer. First, regions of the silicon substrate were implanted with an ion, an electrically charged atom. Afterward, an insulating layer of silicon dioxide was added, followed by a conducting material called polysilicon; it’s poly-si or just poly to its users. Poly was routed over the gaps between the ion implant regions. The ion implants, together with the poly routes above the gap, formed the transistor, the on/off switch. Next, another silicon dioxide layer covered the poly. Finally, a metal layer, typically aluminum or copper, was deposited to route electrical signals around the chip. Vertical interconnects, called vias, between the poly and metal layers were formed using a similar process.

When the wafer fabrication is completed, machines are loaded with programs to test each site. The tester’s robotic arm lands a set of wiry electrical probes on the input/output signals of one of the sites, then runs the test program. Microscopic particles that land randomly while the layers are fab’d cause some sites to fail testing. If the site fails, the tester inks it with a black dot. The tester continues through all sites on the wafer. If high-speed testers, room-sized behemoths, are needed, the amortized cost of the tester time can be a significant increment to the cost of the finished chip.

Let’s assume that a single defect will destroy the chip.¹⁷ If the average number of defects per chip is greater than or equal to one, then the average chip won’t function; many will have one or more defects, but some will have none at all. To illustrate with numbers, if 50 out of 200 chips are OK, the yield is 25%. If the overall wafer cost is $1,000, then each good chip costs $20 at this stage of manufacturing.

If the chip size is halved, the density of defects per chip is also halved and the testing yields more chips without defect, reducing the chip cost compared to that of the bigger chip.¹⁸ Of course, the smaller chip won’t pack as much functionality as the bigger chip. Computer chip sizes, prior to packaging, range from a few square millimeters for a $0.25 microcontroller used to flash the taillights in a car, to hundreds of square millimeters, priced at over $1,000, for the CPU in a powerful engineering computer.

After every site is tested, the wafer is diced