1. Do More Parts Mean Lower Costs? Moore's 1965 Estimate of Yearly Doubling in Circuits
In 1965, Gordon Moore estimated how the "complexity" of circuits would grow in an article for the magazine Electronics [1]. In a circuit with few parts, the cost per part goes down as you add more parts. However, if you pack too many parts in, more products fail to work properly (yield drops), and the cost per unit starts to rise again [1]. This means there is a "right number" of parts where the cost is lowest.
Moore stated that this "right number" had been increasing at a rate of roughly doubling every year. He estimated it would reach 65,000 parts per chip by 1975 [1]. What he was counting here was the complexity where cost is minimized, not just any increase in part count. Note that this was "every year," not "every two years." So, why does an estimate for about 10 years remain a topic of discussion for more than half a century?
2. Moore Called His 1965 Article a Rough Extension, Yet It Predicted Home Computers
In a 2015 interview, Moore is said to have remarked that his 1965 article was just a rough extension, assuming it would continue doubling every year for 10 years [3]. However, that same article contained a passage imagining future life. It suggested integrated circuits would bring "surprises" like home computers (or terminals connected to central calculators), automatic control of cars, and portable personal communication devices [1].
The text only goes as far as "portable personal communication devices," so it cannot be said to predict the smartphone specifically. Still, we can read that Moore envisioned small parts moving beyond research labs and large computers in 1965.
3. The Intel 4004 Had 2,300 Transistors, and Moore's Estimate Was Revised in 1975
On November 15, 1971, Intel released the "4004." It is considered the world's first commercial microprocessor, containing 2,300 transistors. It was manufactured using 10-micrometer technology and operated at about 740 kilohertz. It was developed for a calculator made by the Japanese company Busicom [4].
It is reported that Mr. Masatoshi Shima, an engineer from Busicom, joined Intel's engineers in the development. However, since details of each person's role rely on secondary sources, we simply state here that he "joined" [4]. In 1975, Moore revised his estimate to "doubling every two years." Because the wording of the revision varies slightly depending on the source, we only note that it was changed to "every two years" [2]. The name "Moore's Law" is said to have been given by Carver Mead [2].
4. From 2,300 to 92 Billion: How Many Transistors Are in Apple's M3 Chips?
In October 2023, Apple announced the "M3," "M3 Pro," and "M3 Max." The number of transistors is 25 billion for the M3, 37 billion for the M3 Pro, and 92 billion for the M3 Max [5]. Since the 4004 had 2,300 transistors, the transistor count of the M3 Max is about 40 million times higher (92 billion ÷ 2,300), and the M3 is about 10 million times higher. We have checked the digits carefully to avoid confusing "billion" and "ten thousand."
Doubling 25 times gives about 33.5 million times; doubling 26 times gives about 67 million times. The factor of 40 million falls in between. From 1971 to 2023, which is 52 years, this works out to about 25 doublings, or roughly once every two years. This calculation is an estimate by the author and not a number written in the source materials. So far, these numbers refer to the "count" of transistors. What about the actual "size" of one transistor today?
5. "3 Nanometer" Is Not the Size of a Part but the Name of a Chip Generation
The 4004 was made with 10-micrometer technology, which is 10,000 nanometers. Comparing just the numbers, "3 nm" represents a difference of about 3,300 times (10,000 ÷ 3) [4]. However, today's labels like "5 nm" and "3 nm" do not refer to the actual length of transistor parts. According to the IEEE industry roadmap (IRDS, 2021 edition), for the "5 nm" generation, the gate length is 18 nanometers, the spacing between gates is 51 nanometers, and the spacing of the finest wiring is 30 nanometers [6].
In the past, the number in the generation name represented the gate length. Around 2011, Intel's names began to use numbers smaller than the actual lengths, and now companies use these numbers as names for new generations [7]. For cutting-edge lithography that prints fine patterns, extreme ultraviolet (EUV) light with a wavelength of 13.5 nanometers is used. It is said that a laser hits droplets of molten tin twice to create this light, repeating this process 50,000 times per second [9]. Reading the number in the name separately from these actual dimensions helps avoid confusion.
6. Why Smaller Transistors Stopped Making Chips Faster and Chipmakers Started Adding Cores
In 1974, Robert Dennard showed a prospect that making transistors smaller would allow them to run faster while using less electricity. This idea is called "Dennard scaling" [8]. This prospect stopped holding true around 2005–2007 because power density increased due to leakage current as transistors got smaller. The years are given as a range in the source, so we write "around" [8].
As a result, the clock frequency, which is a measure of operating speed, stopped increasing around 4–6 gigahertz. Chip companies shifted direction toward "multi-core," placing multiple computing cores on one chip [8]. This can be read as moving toward "lining them up and running them simultaneously" rather than "making one faster." Since opinions differ on whether Moore's Law continues or ends, we do not state it definitively here.
7. Check Moore's 1965 Estimate Yourself by Doubling 1 Ten Times
Start with 1 and double it 10 times. You get 1, 2, 4, 8... up to 1,024. Moore's 1965 estimate of "doubling every year for 10 years" was a calculation anticipating about 1,000 times growth.
Next, what happens if you double once every two years for 10 years? That is 5 doublings, resulting in 32 times. Using a calculator to make a table showing where the difference between doubling every year and doubling every two years opens up can help clarify things. Try writing "doubling every year" and "doubling every two years" for 10 years on paper side-by-side. You will see from which year the gap becomes visibly large. If you want to read the original article, Moore's 1965 article is reprinted [1].