6
The Chips Industry: Moore and Rock’s Laws
The electronic components industry predates the computer and connected objects industry that we commonly use. After the Second World War, the invention of the transistor (in 1947) preceded the development of printed circuits by several years. These innovations gave a very important boost to this industry, which essentially formed the technical base for digital services. This industrial production has changed over time: some large companies chose to integrate the design, and sometimes even the manufacture, of the components necessary for their own products (computers, telephones, weapons systems, automatisms, etc.); the facts did not always prove them right. Others deliberately specialized in the production of electronic chips; significant scale and scope effects enabled them to master the technical, commercial and financial risks inherent in this highly capitalized sector. Today’s leaders are among them.
This industry is quite original. It has two main branches: one mainly produces electronic memories and the other produces logic processors, which in turn give way to computers, digital telephones, game consoles and automatisms. The storage capacity of the memory chips and the computing speed of the processors increase according to an empirical law, which is attributed to Moore, one of the founders of the American company Intel which has been dominating this branch since the 1970s. Over the years, the technical progress of semiconductors has been spectacular: the speed and recording capacity of electronic chips almost doubles every other year. The price of components decreases correlatively, in the same percentage and at the same speed as their performance increases. However, this progress is not free of charge, because the investment needed to design and build a production line increases as steadily – or almost as steadily – as the performance of electronic chips.
This chapter, written by an expert in the information industry, describes the economic, managerial and financial consequences of the bronze law that has guided the semiconductor industry since its inception: a lot of risk has to be taken a long time in advance in order to maintain your position over time, and there must be significant margins on current sales to finance huge developments and investments that make it possible, in later years, to produce chips that are more powerful, faster and less expensive than the ones being currently produced so that a demanding, diversified and global clientele can be served1.
Global by its very construction, this industry is absolutely essential to the contemporary world. However, it escapes the patterns of political economy because producers are global giants that are very few in number; markets are global, increasingly diversified and growing rapidly. The enormity of investments imposes wise financial management; the factories rarely employ personnel because the production is entirely automated. Lastly, alliances and cooperation between customers, suppliers, subcontractors and even competitors are progressive and subtle. The careful observers of our time have characterized it as both an industrial and a post-industrial era2. There is no paradox or contradiction here, given that semiconductors constitute the soul of the machines that provide the world population with the facilities of the digital era. An enlightened perspective on this branch of industry is therefore useful for understanding our times3.
The developments discussed in this book are ultimately based on advances in electronic components, which form the ultimate foundation of the information society. Since 1975, we know that these advances follow Moore’s famous law4, which states that the density of integrated circuits on a silicon chip doubles every two years; or its alternative version: the speed of electronic components doubles every two years. Many pundits predicted the end of Moore’s law at the beginning of the 21st Century, because it would have had to come up against both particle physics and the gigantic cost of production lines.
Some words about technique: where do we stand today?
Moore’s law is most often considered to reflect some sort of natural phenomenon. In reality, it is the result of intense research and development activity that constantly requires huge investments. Every year since 1992, and under the aegis of the Semiconductor Industry Association, experts from all companies concerned have produced an extensive document (International Technology Roadmap) which describes the problems to be solved so that Moore’s law can continue to be applied in the next fifteen years. Many of these problems are classified as “difficult” or even as having “no known solution”. Each company then decides, quite independently of the others, to invest itself (or not) in the necessary research, in order to ultimately maintain the miracle that has been going on since 1975: the continued application of Moore’s law, even if the progress actually made has not always been precisely on schedule. This is a fine example of the effectiveness of a spontaneous organization of independent operators, which is in no way because of a coercion or superior authority. Only the common interest of this inter-profession inspires the major annual work of the International Technology Roadmap; companies remain free to pursue their own research, at their own expense and on their own account.
The technology evolves in steps, each defined by the width of the lines engraved in silicon. The smaller the width, the more functions can be placed on the same surface and the faster the circuits are. Towards the end of the last 20th Century, this width was around half a micron (thousandth of a millimeter). Today’s high-end chips are 14 nanometers (millionths of a millimeter), which can accommodate 30 million transistors in a square millimeter and more than 7 billion transistors on a two-centimeter chip side5. The size of a chip cannot however be reduced in the same proportions. For a general-purpose processor to connect with the outside world, it needs a few hundred connectors that cannot fit on much less than two square centimeters. Only memory chips and certain specialized chips, whose function requires fewer points of contact, can accommodate smaller surfaces. It is therefore the number and technology of external connections that dictate the chip’s dimensions and not its content. Conversely, for the same surface area, it is possible to absorb functions hitherto entrusted to other components in a new chip, which is a natural way of using the potential of transistors while increasing the chip’s value and therefore its potential selling price.
It is difficult to imagine what these dimensions represent on a human scale. Let us imagine that a chip is enlarged so as to allow human beings to pass through the circuits instead of electrons. This would require replacing the fourteen nanometer lines with paths that are one meter wide, separated by hedges of comparable width. A typical 2 cm chip would then become a 1,428 km2. The complexity of such circuits is therefore equivalent to that of a labyrinth of corridors that would cover France, the Iberian Peninsula and Germany according to a grid of two by two meters, populated by 30 times more transistors than these countries currently have inhabitants6. It is probable that such electronic chips are the most complex object ever produced by man. This is why their production presents formidable physics problems that are reflected in the high cost of production facilities, where very high-precision equipment that uses cutting-edge technologies must be installed in a very unique environment: extreme atmospheric purity, perfect physical stability of the facilities, extreme precision and perfect safety of fluid circulation and temperature control, etc.
High-tech production
The production cycle begins with the manufacture of an ultra-pure cylindrical silicon crystal made from molten silicon7. Current techniques make it possible to obtain cylinders that are 30 centimeters in diameter and more than two meters long, which is already quite a technological feat. This cylinder is then cut into wafers that are less than half a millimeter thick, on each of which numerous identical chips will be manufactured simultaneously. The industry is currently developing the necessary tools to reach a diameter of 45 cm, making it possible to engrave 2.25 times more chips per wafer. The silicon wafers then endure a succession of chemical and thermal treatments corresponding to the different levels and to the different reagents and substances used in manufacturing. At each step, of which there are about 20 for a microprocessor, there is a pattern that shows the spots where the reagents have been used on the surface of the chip, on which the latter must operate. These patterns are engraved in quartz masks and are transferred to the silicon wafer in multiple copies by a photolithography device, using the extreme ultraviolet produced by a laser8.
This photolithography is repeated with different masks at each stage of chemical and thermal treatment, which requires that the masks and silicon wafers be positioned and repositioned for the next operation with an accuracy that is consistent with the size of the circuits themselves. Manufacturing therefore requires equipment that allows stable positioning with nanometric accuracy. The machines, of which there can be hundreds, are locked up in totally sterile and anti-vibration rooms, built in earthquake-resistant buildings. The transportation of the wafers between the various processing devices requires extremely precise automation. As for the many reagents used, which are generally in gaseous form, they are often toxic and must be free of all impurities. This means that their circulation within the plant requires very safe systems. At the end of the actual manufacturing process, each chip is individually tested in order to mark it and subsequently eliminate those that may be defective, either due to a defect in the silicon crystal or due to an incident during the manufacturing process. The wafer is finally cut into chips; those marked as defective are eliminated, and those that have passed the manufacturing control are fitted with their connectors and embedded in a protective resin.
The environmental constraints that weigh on all these operations are an incentive to avoid as much human presence in the manufacturing area as possible and to thus fully automate the production lines, which presents, among other issues, difficult computer problems. All this translates into huge investments that increase from one generation to the next, as defined by Rock’s law9 (much less known than Moore’s), which states that the production cost would double every four years. This cost was about one billion dollars in 2000. According to Rock’s law, the cost should therefore have risen to 16 billion dollars in 2016 and approach 32 billion dollars in 2020.
An original economy
The nature of the process summarized above requires that the production capacity of factories increase with each generation; this capacity represents a minimum threshold below which it is impossible to design an economically viable factory. On the other hand, the marginal production cost remains roughly proportional to the chip’s surface area, regardless of its circuit content and level of technology. Therefore, producing one more chip only costs a few square centimeters of silicon and a few drops of reagents, or a few dollars. This disproportion between fixed and variable costs calls many of the usual concepts of industrial economics into question. Once the factory is built, its operator has an interest in running it at its maximum capacity, which can amount to tens of millions of chips per month. To do this, the industry will set its selling prices as low as possible in order to sell all of its production, while making a sufficient profit to invest in the next generation. The selling price of an electronic chip is thus an independent variable that the supplier sets according to strategic and commercial considerations that are specific to its situation and its business project, and not according to an accounting cost price.
When selling chips, each manufacturer must ensure that it has a sufficiently high gross margin to cover its investments in order to survive until the next generation. Each chip sold today must therefore bear, in addition to variable and overall costs of business, a share of current and planned investments, which determine a survival price10 for the firm. Given that it is essentially made up of distributed fixed costs, the higher the volumes sold, the lower the survival price, thus reflecting the size advantage. Each manufacturer must therefore strive for the highest possible volumes for each technology, by expanding its catalog, optimizing its sales prices and taking commercial action. Globally speaking, the higher the volumes a supplier provides, the highest its chance of survival!
Each general-purpose microprocessor is typically launched at a price of a thousand dollars or more, depending on the progress achieved over the former chips; this price drops a few years later, below a hundred dollars when the microprocessor somehow becomes obsolete and factories depreciate. Ultimately, neither the price of the various chips at any given time nor the evolution of these prices over time have anything to do with manufacturing costs. That’s why we can now buy plug-in USB memory sticks for 25 euros (XtraPC) and 128GB USB memory cards for 60 euros. A deal that was unimaginable just 20 years ago!
Rock’s law leads to concentration
If the capacity per factory grows faster than the market, then the number of factories and suppliers are doomed to decline. If competition brings prices down to below the survival price of certain companies, they are likely to be eliminated by the time the next generation starts. Rock’s law thus implies that the survival price of each competitor increases with each new generation, which should automatically lead to the gradual elimination of weaker producers. This analysis has led many pundits (including the author of this chapter11) to predict that there would probably only be one chip supplier left by 2015/2020, and that progress may well stop there because manufacturers might be no longer able to make the necessary investments. The misdeeds of Rock’s law would consequently stop the virtuous effects of Moore’s law!
However, it did not go as planned. While the previous analysis predicted that the number of companies and factories would drop to a few units by 2017, there are still just over 200 active manufacturing factories today, belonging to at least 40 companies12. This results from the fact that we overlooked two things: in the race for performance, as the leading company shrinks, it leaves behind factories and equipment in perfect working order whose equipment has been amortized, fitted with technologies from previous generations that may very well suffice for certain applications13. Those who accept a follower role, reduce their risk: most of the 200 existing factories over the world were built before the year 2000 and work at levels higher than 100 nanometers to produce memory chips, processors and various other components. Out of the 200, only seven are at the cutting edge of the latest technology: a very fine 14 nm etching on a 300 mm diameter silicon wafer.
Meanwhile, the market for electronic chips has exploded and diversified considerably. In addition to conventional computers, which still accounted for more than three-quarters of processor and memory outlets at the turn of the century, new applications have emerged, such as game consoles that require high-performance processors, or entry-level telephones that work with simple processors, whereas more powerful smartphones are now genuine portable computers. In 2000, almost all users felt the need for more powerful computers; in 2017, on the other hand, a large majority of computer users (including ourselves) are perfectly satisfied with processors from the 2000s that only cost a few hundred euros, fully suporting an immense variety of applications and situations. The most advanced technological progress in microelectronics is no longer essential for everyone; manufacturers would rather extend the life of their existing factories, and charge fairly low prices in order to encourage their customers to continue with “old” technologies. This evolution has allowed many chip producers to specialize in applications and markets that are compatible with mature technologies.
De-integration, specialization and reconfiguration
Another major phenomenon: in order to ensure a sufficient load, some electronic chip producers have opened their factories to third parties in the form of contract work. In return, some specialize in the design and distribution of components without investing billions of dollars in production lines for which they could not ensure a sufficient load. The electronic components industry has thus embarked on a process of disintegration and reconfiguration that is quite similar to that experienced by the computer industries between 1980 and 2000. Some firms are pure foundries who only work as subcontractors for others on a contractual basis; others are pure designers, known as fabless14 firms such as Broadcomm, Qualcomm, Apple or Nvidia, who design and market chips that are manufactured by foundries; all conceivable combinations between these various activities do exist: design of certain chips, subcontracting of part of the production, etc.
There are about 20 pure foundries in the world, among which there are only four pioneers: the Taiwanese TSMC, the American Globalfoundries and the Chinese UMC and SMIC, who continue to invest in the latest levels of technology to perpetuate Moore’s law. Others specialize in older technologies and dropped from the technology race. Some suppliers of specific components maintain foundries dedicated to their own needs, such as the Korean Hynix and the American Micron, which are both memory suppliers. Large groups such as Toshiba, Sony and Texas Instruments still produce chips that they incorporate into other products; concurrently, they are more or less present on the chip market itself. Other companies specialize in auxiliary processes such as the manufacture of silicon wafers or of lithographic masks or the final packaging of chips; they work on a contractual basis for other companies. Lastly, the two leading firms (the American Intel and the Korean Samsung Semiconductors) have retained an integrated structure; they carry out all the operational characteristics of this industry (from the design to the final packaging of the chips) while offering competitors their foundry services, as do pure foundries. Intel even goes so far as to build its own production tools like steppers.
How to stay in the lead pack?
Consequently, the dynamics of monopoly concentration only concern the few companies that intend to remain at the forefront of the race for the latest generation of technology15. As the movement has been engaged for a few years, this list is short: American Intel and Globalfoundries, Korean Samsung, Taiwanese TSMC, Chinese UMC and SMIC. In 2015, Intel acquired Altera when the company wanted to move to 14 nm etching, and Globalfoundries acquired IBM’s component factories that were at the same level of development.
Along with the efforts made by the industry to maintain Moore’s law, efforts also tend to slow the effects of Rock’s law. Thanks to the progress made in the manufacturing and handling of silicon wafers, the cost of a full production factory, which Rock’s law projected to be around 32 billion dollars in 2017, actually remains at about 14 billion dollars to date. In addition, it is possible to partially reuse existing premises and facilities instead of having to build a new factory for each generation of chips. However, moving from one generation to the next still costs at least five billion dollars, a sum that will probably rise to 10 billion dollars by 2020. Only Intel and Samsung Semiconductors, which each have an annual turnover of around 60 billion dollars (and, at the very least, TSMC with a turnover of around 25 billion dollars per year), will have the capacity to make such investments. The other manufacturers, of which the annual sales are between three billion for SMIC and six billion for Globalfoundries and UMC, will at some point either have to give up the race and become secondary foundries or merge with a better-placed champion, or else they will run to ruin with their efforts16. The recent cooperation between Samsung, Globalfoundries and IBM’s research laboratories to reduce the line width to 5 nm could be the harbinger of a future industrial concentration.
At the rear of the pack, SMIC will probably not be able to continue the race in isolation. This Chinese semiconductor manufacturer should logically either give up or merge with its compatriot UMC, which will only postpone the problem for two or three years. It would therefore be more rational (at least in industrial terms) to go straight to the next stage, but this would preclude a possible government intervention that could have other motivations than mere industrial logic. In a few years, Intel and Samsung may find themselves confronting each other. It should then be remembered that Intel mainly focused on computers, including very powerful machines, while Samsung mainly targeted phones and tablets. Moreover, while electronic components are basically Intel’s only business, Samsung Semiconductors is only part of a very diverse chaebol. Faced with a major obstacle, Samsung could therefore simply resort to abandoning the race for performance and join the group of companies who manufacture components17.
For the industry as a whole, the consequences would certainly slow down or even stop the race for miniaturization and the speed of circuits. However, the effect on industry in general may be marginal, as there are already powerful chips that are cheap enough, miniaturized enough, efficient enough energywise and diversified enough to be incorporated into countless objects, including clothing and even into a human or animal body (for example, for medical purposes). Additionally, their share in the cost of the devices and services in which they are incorporated is rarely predominant.
Some limits have already been reached. The frequency of circuits has been blocked at a few gigahertz since 2004, due to heat dissipation; the speed improvements are now based on the increasing parallelism of treatments and the complexity of the circuits. The transition to 450 mm diameter silicon wafers has been delayed. The next physical limit to be crossed will be to further reduce the width of the lines engraved on silicon; but below 5 nm, quantum phenomena disturb signals. This limit could however be circumvented by stacking the circuits on several layers, just as it is already done for memory chips which comprise several tens of layers.
There is therefore still room to produce even more powerful circuits, notably through functional enrichment and the migration of certain functions from software to hardware, which would lead to more and more “intelligent” chips for the most advanced applications. That being said, caution must still be applied: I would only be half surprised if today’s prognosis turned out to be as bad as the one I made in 1996!
Bibliography
[ANY 16] ANYSILICON, Top-20 Semiconductor Suppliers 2016, available at: http://anysilicon.com/top-20-semiconductor-suppliers-2016/, 2016.
[DRÉ 96] DRÉAN G., L’industrie informatique, structure, économie, perspectives, Masson, Paris, 1996.
[ITR 15] ITRS, International Technology Roadmap for Semiconductors 2.0, Report, available at: https://www.semi-conductors.org/clientuploads/Research_Technology/ITRS/2015/0_2015%20ITRS%202.0%20Executive%20Report%20(1).pdf, 2015.
Refer as well to the major firms’ websites: Global Foundries, Hynix, IBM, Intel, Samsung, TSMC, Texas Instruments, etc. Wikipedia articles on Moore’s Law and semiconductor manufacturing may also be useful sources: https://en.wikipedia.org/wiki/List_of_semiconductor_fabrication_plants.
Chapter written by Gérard DRÉAN.