5The forecast that aged

6Empirical CPU doubling time across the dataset is 72.03 years. 8Moore's 1975 revised prediction was 92.00 years. 10Forty-eight years, seven orders of magnitude, 170 chips — and the slope deviates by 0.03.

14In numbers: the Intel 4004 in 1971 carried 2,250 transistors. 15The AMD EPYC Rome in 2019 carried 32 billion. A 1614.2-million-fold 17climb, on a curve so steady that a pencil line drawn in 1975 still threads through the points half a century later.

18None of this was law in the physical-constants sense. 19Gordon Moore wrote a four-page article for Electronics 20magazine in April 1965 and drew a straight line through five data points. 21He projected that "by 1975 economics may dictate squeezing as many as 65,000 components on a single silicon chip." The piece had no formal model. 22The exponent has held anyway.

28The man and the curve

29When the article was published Moore was Director of R&D at Fairchild Semiconductor; he hadn't yet co-founded Intel. 30The phrase "Moore's Law" wasn't coined for another five years. 31In 1975, after a decade of new data, he revised the doubling rate from 12 months to 24 — the version that has stuck.

4

5

36The first chip on the curve, six years after the prediction, was the Intel 4004: 2,300 transistors at a 10-micron process node, designed by Federico Faggin's team for Busicom's printing calculator. It is the anchor point of the entire CPU dataset — the leftmost dot on every Moore's-Law plot ever published.

38Three families, three slopes

39Plot all 326 chips on one chart and three roughly parallel lines emerge — but they are not the same line. CPUs double every 402.03 years, GPUs every 411.85 years, RAM every 421.56 years.

43The differences are mechanical, not magical. 44RAM is mostly bit cells; doubling capacity directly doubles transistors, so RAM tracks lithography almost one-to-one. 45CPUs spend their transistor budget on logic complexity, which has historically scaled at exactly the rate Moore drew. GPUs sit in between — until the late 2000s, when parallel arithmetic became the default tool for both graphics and machine learning, and GPU counts started catching CPUs.

By 2017–2018 the families converge: a 21-billion-transistor GPU (Volta), a 23-billion-transistor CPU (GraphCore), and a 137-billion-transistor RAM module are all sitting in roughly the same band on the chart. 48Before the late 2010s no such band existed.

51Where Moore was wrong

8

52The 1965 paper aimed at 65,000 components by 1975. 53The CPUs that actually shipped in 1975 topped out at 545,000 transistors. 55The 1976 maximum was 8,500. Moore overshot by a clean order of magnitude — exactly the kind of error any forecaster makes when extrapolating a young exponential.

62In 1975 he halved the doubling rate from 12 months to 24, and the revised line aged into a law. The version most people know — Moore's Law as a steady two-year doubling — is not what the original article said. 64It is what Moore said after 65checking against a decade of evidence.

66Where it slowed without anyone noticing

Split the CPU record at 2005 — the year Dennard scaling collapsed and clock speeds plateaued — and the doubling time stretches: 682.02 years before, 692.51 years after. 70CAGR drops from 40.9% to 31.8%. 71GPUs show the same break, even more sharply: 1.62-year doubling before 2005, 2.53 years afterward.

72This slowdown is the consumer reality that almost no Moore's-Law graph shows. 73After ~2005, transistors kept doubling, but they stopped translating into proportionally faster single-threaded performance. Clock speeds froze around 3–4 GHz. 75The new transistor budget went into multi-core, GPUs, and AI accelerators — parallelism instead of frequency. 76The "exponential is still going" story is true; the "your laptop will be twice as fast every two years" story is not.

79Crossings, jumps, and the chips that beat the line

80The dataset crosses each power-of-ten transistor milestone roughly every seven years. 81Past 1,000: Intel 4004, 1971. 82Past 10,000: Intel 8086, 1978. 83Past 1 million: Intel i860, 1989. 84Past 1 billion: dual-core Itanium 2, 2006. 85Past 10 billion: 32-core SPARC M7, 2015. Seven 10x crossings in 44 years — exactly the rhythm a 2-year doubling produces.

87The crossings are not smooth. 88The single biggest year-over-year jump in the running CPU max came in 1982, when the Intel 80286 leapt 8911.65x 90past the previous high of 11,500 transistors. 91The Motorola 68020 in 1984 jumped 8.64x. 92Dual-core Itanium 2 in 2006 jumped 6.80x. 93The exponential advances in shoves, not glides, and a few well-placed chips do most of the work.

94Residuals to the per-family fit reveal two families of outlier. 95Above the line: Intel's Itanium 2 series (the dual-core Itanium 2 sits +0.95 log-units above its year's predicted count, the largest residual in the CPU record), and SPARC64. 96Below the line: ARM 9TDMI, ARM 1, ARM 2, ARM Cortex-A9, Atom — chips designed to trade transistor count for power efficiency. 97The same line that traces the law also separates two design philosophies: maximalist server logic on one side, low-power embedded chips on the other.

101The Moore line on the data

102Drawing Moore's 2-year line through the Intel 4004 at 1971 and projecting it to 2019 gives a predicted 10338.6 billion transistors. 104The actual maximum CPU in 2019 was 32 billion. Moore's prediction overshoots by 21% — within the dataset's own stated 10–20% accuracy band.

108On a single log-scale chart of all 326 named chips, the Moore line passes through the cloud almost like a regression. 109Forty-eight years apart, the prediction line and the empirical chips meet at the same place — to within one order of magnitude — every year. 110There is essentially no other public, multi-decade forecast in technology that has held this well. 111That is the headline finding.

After the dataset

112The International Technology Roadmap for Semiconductors — the global document that effectively codified Moore's Law — released its final report in 2016 and dissolved. 113The closing report concluded that traditional 2-D scaling would hit an economic wall by 2021, after which the industry would lean on 3-D stacking, advanced packaging, and specialised accelerators rather than simple lithographic shrink.

114The dataset cuts off in 2019, so it captures the very first chips of the post-ITRS era — Apple's A12X, Nvidia Turing, AMD EPYC Rome — without yet reaching the trillion-transistor wafer-scale chips that arrived afterward. 115Apple's M1 Ultra in 2022 fused two M1 Max dies on a single interposer for 114 billion transistors. 116Cerebras's WSE-3 in 2024 is a single piece of silicon with 4 trillion transistors — more than 100 times the count of any chip in this dataset.

123Whether the line continues, bends, or transforms is the open question. Inside the 1971–2019 window, though, the answer is settled: the most famous quantitative prediction in modern technology was, on the data, almost exactly right.