Predicting semiconductor business trends (Dr. Walden Rhines)

• Tools for predicting semiconductor trends
  • Learning curves
    • Cost per transistor vs. total number of transistors manufactured so far follows a straight line on a log/log graph. This trend was already in place since vacuum tubes (with more or less the same slope) and the general principle is also true for many other items.
    • When such curves are predicting an impossible situation in the future, something will have to change (e.g. as some point price of testing per transistor was decreasing slower than total price of manufacturing per transistor and that eventually changed because of compression -- something had to give).
  • Known constants
    • Price per wafer (of constant area) is constant (around $32.5/sq.in.).
    • EDA cost per transistor is on a log/log line parallel to the total cost per transistor. The same is true for fabrication equipment, mask making equipment, assembly equipment, testing equipment.
    • Total semiconductors shipped by area are growing linearly at ~0.27B sq.in. per year. Total IC revenue is also growing linearly. OTOH, IC units sold are growing exponentially (which seems to imply that chips are becoming smaller in area).
  • Moore's law
    • Formulated in transistor density / year it's not sustainably true, but it's true in cost per transistor / total number produced, because learning curves, and that's enough. We should not worry.
  • Gompertz curves
    • S-curves can be useful for modelling many different types of technological progress (and other processes in nature). The formula is: y = a * exp(b * exp(-c * t)).
    • The derivation of this formula based on a mathematical model of a market can be seen in this paper.
    • Forecast of laptop manufacturing numbers based on Gompertz curve made in 2001 was still quite accurate in 2016 (after the point of maximal growth was already passed).
    • For silicone transistors the point of maximal growth is predicted for 2038.
• Competitive dynamics in semiconductor industry
  • Consolidation
    • The basic model is that in the early days of an industry there's a lot of innovation and new companies are created often. After the industry matures the focus is on efficiency and economies of scale help the big players acquire the smaller ones.
    • Not clear if this is actually generally true. The steel industry, on which this analysis was originally based is currently not as consolidated as in the past. Automotive has very high degree of consolidation but now it's less and with emergence of electric cars it is likely to further reduce.
    • The semiconductor industry so far has been deconsolidating. Also the leaders changed many times.
  • Specialization
    • Companies often increased their profitability by specializing (e.g. TI).
    • R&D investment is at ~14% of revenue. This would not make any sense in a stagnating industry where you win by efficiency and economies of scale.
    • It seems that there's still a lot of innovation in semiconductors.
  • EDA segment - it's quite consolidated at the moment with Mentor, Synopsis and Cadence holding 85% market share. The leaders changed over time though.
• Most profitable (in terms of gross profit margin) segments of semiconductor business are FPGAs (66%), Analog components (52%) and microprocessors (56%). Those segments have more differentiation and higher switching costs.
• Japanese are optimizers, not inventors. When standards are stable, Japan will do well. When standards are evolving, not so well.
• After the events of the late teens, China sees semiconductor independence (from the US) as a strategic goal. The rest of the world also took note, so countries are more interested in developing a local semiconductor supply chain.
• Future:
  • AI is likely to drive innovation in custom silicone.
  • Edge computing, IoT will create a lot of demand for chips, although the margins are likely to be low there.
  • 5G will require lots of new base stations.
  • Automotive applications are growing will continue growing, especially with achievement of self-driving capabilities.
  • Makimoto's Wave (oscilation between standardized and customized solutions) is moving towards customized now.
  • Silicon transistors will continue being the foundation of compute for another 20-ish years.