Artificial Intelligence in Science
The end of Moore’s Law? Innovation in computer systems continues at a high pace
Introduction
Are ideas to improve computing systems getting harder to find? A key metric tracking the dramatic progress of electronic technology (denoted Moore’s Law) suggests that progress is nearing its end. This has stimulated fears of a serious decline in the pace of innovation in electronics. Such a decline could have many ramifications; microelectronics are central to practically all industrial products and systems – from kitchen appliances to power generators. However, while the ability to shrink transistors is reaching physical limits, fears of stagnation or decline in the power of computing systems are premature. As this essay discusses, other innovations – additional to those tracked by Moore’s Law – continue to improve the productivity and processing power of computing systems.
Measuring and predicting the progress of a technology-driven field with a single metric can generate inadequate results. Eventually, technologies that advance along established lines of innovation reach a point of diminishing returns. However, rates of progress can then increase because of unanticipated innovations that shift established sources of development. Technical developments in electronics are a great example of this (Kressel and Lento, 2007).
While Moore’s Law has been useful, no reliable and general metric of progress is available because computing systems range so greatly in scale and functionality.
A short description of computing systems
A short review of what makes computing systems function is needed to understand the shifting sources of innovation in electronics technology. The transistor has enabled the digital electronic world. Invented in 1946, it has developed in current amplifiers, storage elements and switches. Interconnected tiny transistors form the integrated circuit chips that underpin computing systems.
The computing power of a system is a function of the available transistor capacity, the speed of transistor switching (i.e. the opening and closing of a circuit), memory volume and interconnection speed. A smartphone today has more computing capacity than a typical mainframe computer in the 1960s.
Moore’s Law tracks the progress of the transistors in integrated circuit chips (also called “integrated circuits”). Such chips perform different functions in computing systems – either signal processing, data storage or combinations of both. The first integrated circuits, built in 1962 at Intel, incorporated only a few transistors. Today, as many as 16 billion transistors can be interconnected on a chip not much larger than a thumbnail. Minimum transistor feature sizes on chips have declined from about 30 microns to about 5 nanometres (for reference, a human hair is around 100 microns thick). In 1960, a single transistor sold for USD 1.00.
All of the above developments mean that computing capacity per unit cost has increased enormously. Today, an integrated circuit chip with 1 billion transistors costs under USD 3.00. Technological wonders had to be invented to achieve such reductions in scale and cost, all while maintaining extraordinary reliability.
The slowing of Moore’s Law
Moore’s Law – an observation rather than a physical law – has held for about five decades. It posits that transistor chip density doubles roughly every two years, with a corresponding decline in unit transistor cost. With declining size, transistor switching speed also improves as power dissipation declines.
As discussed below, progress in increasing transistor density in keeping with the historical pattern is reaching an end. Hence, improvements in electronic systems driven solely by shrinking transistors are at risk. This has led to fears of an end to innovation in computing systems.
However, other innovations continue to improve the economic and technical performance of electronic systems. Good ideas are not running out. Nor is there evidence of declining interest in such research. The key technical issues – much simplified – are summarised in what follows.
Physical limits impact transistor scaling due to the relationship of gate width, transistor performance and photolithography. Gates control the flow of current in a circuit. As active gate width shrinks, it becomes a challenge to maintain transistor performance. It is also challenging to create films of materials at very small dimensions on silicon (using a process known as photolithography).
Below a certain gate width, which now approaches near atomic dimensions, the switching properties of the conventional transistor structure deteriorate. Furthermore, as the problem of patterning becomes increasingly severe, special ultraviolet laser light sources, multiple exposures and extraordinary control of the photolithographic equipment are required. Together, they greatly raise cost.
As a result, unit production costs for transistors start to rise, as contrasted with the decline described as Moore’s Law. Accordingly, today, the most advanced, smallest-feature chips are justified economically only by applications requiring the highest logic and memory performance. This limits design decisions to products expected to sell in high volumes, such as smartphones or large cloud computing systems.
It is important to note that chips represent a declining share of overall system cost (except for several memory chips). This is because the costs of software and peripheral hardware, but particularly software, are rising as a share of system cost. The functional performance of the chips is a more important consideration; in this, there is great progress.
Where innovations are driving progress in electronic systems
Improving chip performance by shrinking the size of transistors is clearly reaching an end. However, electronic system performance is improving due to a number of innovative approaches.
Three-dimensional architecture
New three-dimensional structures are extending transistor performance, while shrinking some of its dimensions. These three-dimensional architectures make better use of the chip area. This is accomplished as follows. All chips incorporate dozens of thin layers (stacks) of different materials. Tailoring the layers in new ways can increase switching speed and lower power dissipation. Furthermore, within the stacks, chips can include sophisticated interconnections of memory and logic elements to increase interconnection speed. Industry sources believe this approach will double the operating chip transistor density over the next decade.
However, progress will be costly because new vertical architectures are hard to manufacture. New state-of-the-art chip production facilities cost more than USD 10 billion each and require a highly skilled workforce. Only a few plants in the world can produce such chips in volume (current manufacturers include TSMC, Samsung and Intel, Micron, SK Hynix and Western Digital).
Integration
Another area of innovation, which can also reduce costs, is in packaging chips to bring logic processing functions, memory and external communications closer together. Packages have been developed where optical fibre and lasers are close to the chip output. In this way, self-contained sub-systems can be integrated economically into a larger system. For example, a new company, AyerLabs, has developed modules that replace internal copper interconnections with optical links. These reduce power dissipation and allow faster inter-chip communications. Novel packaging technology makes it cheaper to build systems compared to assembling individual chips on a circuit board.
These innovations aim to continue to reduce computing costs as data volumes mushroom. In addition, cloud computing is enabling ever-more powerful computing power at reduced cost. The emergence of cloud-centralised computing capacities supports massive computing needs in a cost-effective manner and on an as-needed basis. With cloud technology from Google, Amazon and other providers, enterprises can muster the computing power needed from organisations that have built (and continue to build) large-scale, cost-effective computing centres.
Finally, it is highly likely that quantum computing systems using technologies different from classical computers will one day become practical for large-scale computing systems. This will raise computing capabilities to new heights of performance. Worldwide research is attempting to solve great engineering challenges facing developers of quantum computers. Laboratories in the United States and elsewhere are making good progress and providing practical demonstrations with small systems.
Alternative metrics to changes in transistor density
Given the diversity in types, scales and functionalities of electronic systems, no reliable general metrics of progress in computing systems (or even integrated circuits) exist yet. However, various attempts have been made to find metrics. In IEEE Spectrum, Moore (2020) described approaches to develop a useful metric for monitoring progress in the field. These approaches combined measures of progress in changing chip parameters such as interconnections. However, these attempts are hindered by changes in system performance occurring along many parameters. The development of metrics that define performance of specific systems, such as smartphones or cloud computing services, is more likely.
The ongoing debate about the value of imperfect metrics compared to none at all is healthy and invigorating. A field that appears to have reached its limits will not attract the best students. Moreover, the question of metrics may attract great students to a career in microelectronics. It is important, then, that the reality of continued progress, and the corresponding opportunities, is widely understood.
Which parts of the world will deliver the needed innovations?
All major chip manufacturers invest substantially in advanced product development. However, the fruits of this work are not always published. Published results tend to come from academic institutions and government-funded laboratories.
In the past, when the United States dominated semiconductor manufacture, it produced most semiconductor process innovations. Intel was the clear leader. Extensive academic research in places like the University of California, funded by government, was an important source of innovations that found their way into industry. However, offshoring of the semiconductor industry has changed the innovation landscape, as Korea and Chinese Tapei have developed companies with equivalent or possibly superior capabilities. Meanwhile, the People’s Republic of China has also made significant advances (Badaroglu and Gargini, 2021).
In addition, as previously noted, no innovation described in this paper can reach the market without production equipment of increasing sophistication and cost. Only a few companies in Europe or North America are left in this market. Applied Materials, in the United States, is a clear leader in process equipment, along with Lam Research. In advanced photolithography, a Dutch company, ASML, has a near monopoly. These companies maintain high levels of research to maintain their positions and profitability.
Conclusion
For 50 years, the world has benefited from an extraordinary level of innovation in electronics. The ability to scale and manufacture transistors at ever-decreasing unit cost has been a key enabler. However, falling unit cost turns out to have been a simplistic measure of innovation in this field. Industry sources predict a doubling of chip transistor density over the next ten years, not the two years described by Moore’s Law. This does not mean the end of innovation in electronic systems based on semiconductors.
There are many creative ideas for development. As one reason for optimism, innovations have moved from a focus on chip transistor density to the elimination of bottlenecks in system performance. To that end, they have reduced “parasitic losses” (e.g. those related to peripheral capacitance), and decreased inter-system signal delays that reduce processing speed. Furthermore, new transistor and chip architectures extend switching performance limits as device dimensions get closer to the atomic scale.
References
Badaroglu, A. and P.A. Gargini (2021), “System and high volume manufacturing driven more Moore scaling roadmap”, IEEE Electron Society Newsletter, January, Vol. 28, pp. 1-9, www.ieee.org/ns/periodicals/EDS/EDS-JANUARY-2021-HTML-V5/InnerFiles/LandPage.html.
Bloom, N. et al. (2020), “Are ideas getting harder to find,” American Economic Review, Vol. 110/4, April, pp. 1104-44, www.aeaweb.org/articles?id=10.1257/aer.20180338.
Kressel, H. and T.V. Lento (2007), Competing for the Future: How Digital Innovations are Changing the World, Cambridge University Press, Cambridge.
Moore, S.K. (2020), “A better way to measure progress in semiconductors”, IEEE Spectrum, 21 July, https://spectrum-ieee.org/a-better-way-to-measure-progress in semiconductors.