Jorge Quitério Search
Technology

Compound growth in technology improvement

Moore's law is usually presented as a statement about chips. More generally it describes compound growth, and the arithmetic of compounding explains why small differences in improvement rates lead to very different outcomes. The article also shows how uncertain a rate estimated from few observations is.

Consider two technologies. Technology A improves by 5% a year. Technology B improves by 10% a year. After thirty years, how far ahead is B?

A common intuitive answer is "about twice as far". The correct answer is about four times as far: A has improved fourfold and B seventeenfold. Intuition tends to work additively, whereas compound growth is multiplicative.

Savings follow the same arithmetic. Over thirty years, a pension fund returning 7% a year grows 7.6-fold; with an annual fee of one percentage point, it grows 5.7-fold. A fee that sounds small removes about a quarter of the final sum. Rates of technological improvement compound in exactly the same way.

Moore's law is usually told as a story about transistors. In MIT's Management of Technology programme it was presented more generally: many technologies improve their figure of merit (FOM), the number that captures what users care about, by a roughly constant percentage each year. That is compound growth,

FOM(t)=FOM0 (1+r)t, \text{FOM}(t) = \text{FOM}_0\,(1+r)^{t},

This article examines what the formula implies and how difficult it is to estimate rr reliably.

The relevant literature is brief.

Moore (1965) observed that the number of components per integrated circuit had doubled every year and predicted it would continue; in 1975 he revised the pace to a doubling about every two years. The observation became an industry roadmap and, to a degree, a self-fulfilling prediction.

Nagy and colleagues (2013) tested competing laws of progress on 62 technologies and found that exponential improvement in time (Moore's form) and improvement with cumulative production (Wright's learning curve) predict about equally well, because production itself tends to grow exponentially.

Magee and colleagues (2016) collected long series of performance data for 28 technological domains and found that improvement rates are roughly constant within a domain over long periods, but differ greatly between domains, from a few percent a year to over 30%.

de Weck (2022) builds a whole discipline of technology roadmapping on figures of merit and their rates of improvement, and is the source of the aviation example below.

Compound growth over thirty years

Over thirty years, the annual rate makes an enormous difference:

Annual improvement After 30 years
2.5% about 2×
5% about 4×
10% about 17×
20% about 240×

Year to year, all four rates appear as steady progress. Over thirty years, the outcomes range from modest improvement to transformation.

A useful shortcut is the rule of 70: a quantity growing at rr percent per year doubles in about 70/r70/r years. At 5% that is 14 years; at 10%, 7 years. (The 70 comes from ln⁡2≈0.693\ln 2 \approx 0.693.)

The rate implied by Moore's law

Doubling every two years means (1+r)2=2(1+r)^2 = 2, so

r=2−1≈41% per year. r = \sqrt{2} - 1 \approx 41\% \text{ per year}.

The course material put the long-run rate for computing at about 37% a year, close to that figure. Very few technologies sustain anything like it. In general, technologies that process information improve much faster than those that transform matter or energy, a pattern Magee's data shows clearly.

Example: air transport

The course compared the Douglas DC-3 of the 1930s with the Airbus A350-900ULR, 82 years later, on a productivity measure: revenue passenger-kilometres a single aircraft can deliver per flight. It put the ratio at about 100, which means

(1+r)82≈100⇒r=1001/82−1≈5.8% per year. (1+r)^{82} \approx 100 \quad\Rightarrow\quad r = 100^{1/82} - 1 \approx 5.8\% \text{ per year}.

Aviation is not usually described as a Moore's-law industry, yet a rate below 6% sustained for eight decades produced a hundredfold improvement.

Crossover between technologies

The rate matters more than today's performance. A technology ten times worse than the incumbent but improving at 30% a year, against the incumbent's 5%, closes the gap when (1.30/1.05)t=10(1.30/1.05)^t = 10, which is after about 11 years. Incumbents that compare current performance rather than rates of improvement tend to underestimate this.

Estimating the rate from data

The rate is estimated from history. Because the model is exponential, take logarithms,

ln⁡FOM(t)=ln⁡FOM0+tln⁡(1+r), \ln \text{FOM}(t) = \ln \text{FOM}_0 + t \ln(1+r),

and fit a straight line by least squares; the slope estimates ln⁡(1+r)\ln(1+r). On a logarithmic scale a constant rate is a straight line, which makes departures easy to see.

Two cautions apply. A line fitted through three points is highly uncertain, and the figure below shows how wide the resulting interval is. And a perfect fit to the past says nothing about the future if the technology is nearing a physical limit, which is the subject of figures of merit and their theoretical limits.

The figure below illustrates three points: doubling time is about 70/r70/r, small differences in rate become large over decades, and a rate estimated from a handful of noisy points is much less certain than it looks.

Suggested settings:

  1. DC-3 to A350. Press the preset. At 5.8% for 82 years the factor is about 100, and the doubling time is about 12 years, matching the rule of 70.
  2. Moore's law. Press the second preset: a 41% rate doubles in exactly 2 years, and thirty years give a factor in the tens of thousands. Here the rule of 70 says 1.7 years, a reminder that it is a small-rate approximation.
  3. Fewer data points. Set the data points to 3 and press "New noisy data" several times. The 95% interval for the rate varies widely. Now use 20 points. With the same noise, the interval is much narrower: the uncertainty of the slope falls roughly like 1/n1/\sqrt{n}.
  4. Shorter history. Keep 5 points but cut the years to 10. The interval widens again: a slope is easier to measure over a long baseline.

Where the rule of 70 comes from. The exact doubling time is T=ln⁡2/ln⁡(1+r)T = \ln 2 / \ln(1+r). For small rr, ln⁡(1+r)≈r\ln(1+r) \approx r, so T≈0.693/rT \approx 0.693/r, or 69.369.3 divided by the rate in percent; 70 is rounded for mental arithmetic. For the estimate, ordinary least squares gives the slope's standard error as s/∑(ti−tˉ)2s/\sqrt{\sum (t_i - \bar t)^2}, where ss is the residual standard deviation. Fewer points or a shorter span shrink the denominator, and the interval uses the tt distribution with n−2n - 2 degrees of freedom, which is very wide when nn is 3 or 4.

Small differences in rate become large differences in outcome. For technology management this has three practical consequences:

  • Compare rates, not levels. The challenger that is worse today but improving faster has a crossover date that can be estimated.
  • Ask what rate a roadmap assumes, and whether anything similar has ever sustained it. A plan that needs 40% a year from a technology that transforms matter should be examined critically.
  • Report the uncertainty of the rate. A rate fitted to five points comes with an interval, and the interval should be reported with it.

References

  1. G. E. Moore, "Cramming more components onto integrated circuits", Electronics, 38(8), 1965.
  2. G. E. Moore, "Progress in digital integrated electronics", IEEE International Electron Devices Meeting, 1975.
  3. B. Nagy, J. D. Farmer, Q. M. Bui and J. E. Trancik, "Statistical basis for predicting technological progress", PLoS ONE, 8(2), e52669, 2013.
  4. C. L. Magee, S. Basnet, J. L. Funk and C. L. Benson, "Quantitative empirical trends in technical performance", Technological Forecasting and Social Change, 104, 237-246, 2016.
  5. O. L. de Weck, Technology Roadmapping and Development: A Quantitative Approach to the Management of Technology, Springer, 2022.