Source: Content compiled by Semiconductor Industry Watch (ID: icbank) from jbpress, thanks.
Some 60 years have passed since Gordon Moore, one of the founders of Intel, proclaimed Moore's Law. For about 50 years, the number of transistors on semiconductor integrated circuits has continued to grow exponentially in accordance with Moore's Law.
In recent years, Moore's Law has reached its limits. However, even when Moore's Law reaches its limits, the increase in the density of integrated circuits cannot be stopped. The transformation of society to digital, the spread of the Internet of Things and the emergence of artificial intelligence. All of this requires highly integrated semiconductors.
How to continuously improve the integration of integrated circuits? How will the semiconductor industry change? What will be important for the semiconductor industry in the future? Tadahiro Kuroda, professor at the Graduate School of the University of Tokyo, was interviewed for this article.
Q: We've heard a lot in recent years that "Moore's Law is coming to an end." To what extent did Gordon Moore foresee that Moore's Law might reach its limits?
Mr. Tadahiro Kuroda (hereinafter referred to as Kuroda) : In a speech at an international conference in 2003, Mr. Moore asserted that "[the increase in the number of transistors on an integrated circuit] will not increase exponentially forever." However, he added that it was possible to delay the end of exponential growth.
Delay the end of exponential growth in the number of transistors in integrated circuits. I believe this is the role that researchers and engineers engaged in semiconductor research today should play.
Q: You believe that society will evolve from a capital-intensive, industrialized society to a knowledge-intensive, knowledge-based society. Tell us what a knowledge society is and what role semiconductors will play in it.
Haruhiko Kuroda: Have you heard the terms "from things to services" and "from goods to services" recently? In this world, value is not found in material "things," but in "things" and "services," in other words, knowledge. It was a knowledge-based society.
Until now, in industrialized societies, semiconductors were just components. Just like screws and nails, consumers don't know it's being used in the final product. It's the presence of the shadow. Standardized parts need to be available in large quantities at low prices.
In a knowledge society, on the other hand, data is key. By using machine learning to process large amounts of data, various services using artificial intelligence are being implemented in society.
The Internet of Things needs to collect data. The Internet of Things uses devices with built-in semiconductors, such as sensors and communication terminals. In addition, without the semiconductor, the collected data cannot be analyzed.
Devices such as smartphones are used when providing services to consumers that use data analytics. Smartphones are full of semiconductors.
In the knowledge society, semiconductors are everywhere. By then, semiconductors will no longer be limited to the parts industry. Semiconductors can become the infrastructure of a knowledge society.
Q: Contrary to Moore's Law, which drives two-dimensional miniaturization, you mention the importance of "beyond Moore" for the three-dimensional integration of chips. Why do we need to integrate the chip in 3D now?
Haruhiko Kuroda: Since the process node is around 40nm, the price of transistors no longer drops even with further miniaturization.
I think it's easier to understand if you compare it to land in a big city. In big cities, land prices are too high to build even a small bungalow. As a result, buildings are getting taller and taller.
The same is true of semiconductors. While miniaturization itself is possible, it does not reduce costs. There is a need for efficient use of land, i.e. area. Like skyscrapers, wouldn't it be better to stack integrated circuits vertically up and up? Such thoughts are natural.
In addition, in the semiconductor industry, high-rise buildings of 170 meters and 50 stories in the past will need to be increased to 100 or even 1,000 stories while maintaining the same height. This will usher in an era of fierce competition for Z-axis miniaturization.
Q: You have mentioned several times that "agility" (fast, flexible) is an important point in the future development of semiconductor chips.
Haruhiko Kuroda: Until now, the semiconductor industry has focused on cost. But those days are coming to an end.
The next axis of competition is performance and time. By "time," I mean "How quickly can we get this product to market?"
As of 2023, artificial intelligence continues to accelerate. In just over a month since ChatGPT launched, the number of users has surpassed 100 million. In times like these, even if you spend two years creating the "latest" semiconductor, it's still the "latest" from two years ago.
That's why we need agility. Agility is the new value of semiconductors.
In the software industry, time is already a competitive axis. Develop quickly, release to the market, and upgrade based on user feedback. I believe a similar approach will be needed for future semiconductors.
Q: You believe that semiconductor demand will shift from general-purpose chips to specialized chips as an industrialized society shifts to a knowledge society. Why is the demand for specialized chips increasing in the knowledge society?
Haruhiko Kuroda: The bottom line is that dedicated chips are energy efficient.
Energy is a major issue in the semiconductor industry right now. Ai semiconductors compute large amounts of data at high speed. At this point, a lot of energy is consumed.
General-purpose chips are designed so that anyone can use them for any purpose. It is flexible, but not yet optimized.
Semiconductor users with a clear goal of "I want to do something with AI" want a dedicated chip optimized for that purpose. Dedicated chips eliminate unnecessary performance and therefore have high energy efficiency and target domain specific performance.
Since the late 2010s, major Silicon Valley tech companies such as GAFAM and Tesla have started developing their own specialized chips. Previous chip users are now developing specialized chips for their own use.
It takes two years to develop a dedicated chip, with about 200 engineers working together. The investment alone amounts to about 50 billion yen. If it is further put into production, it will cost about 5 billion yen, 10 billion yen. Only large technology companies with deep capital can acquire specialized chips.
There are a lot of people who think, "I want to do this business or use AI to provide services." These people simply don't have the funds to buy specialized chips.
We will provide them with specialized chips in small batches and short lead times, and we will continue to implement new ideas into society. If we can do this, new businesses and services will emerge at the top of society.
Rapidus was established in Japan to address these needs.
Chip development used to be on a 100-point scale, and now it should be on an 80-point scale. If you want 80 points, you can do it in a fifth of the time. Chip development can quickly reach 80 points, but trying to get to 100 points from there requires a lot of time and cost.
The most time-consuming and costly process for manufacturing semiconductors is chip development. By implementing an 80-point chip development process, time and cost can be significantly reduced.
However, there may be trade-offs, such as larger chips and higher manufacturing costs. But given the total cost, it could be cheaper.
Q: In the semiconductor industry, where mass production is the norm, I find it difficult to make a profit no matter how much development costs are reduced.
Haruhiko Kuroda: As you said, the idea of producing semiconductors in small batches in an agile way seems to be difficult for the industry to accept. However, I have recently found that some users agree with me very much.
I'm a researcher.
Studying the world is very rigorous. If one day you fall behind the rest of the world, your research will never see the light of day.
Astronomers, earth scientists, economists and others all compete in academic fields that deal with large amounts of data. The challenge is how fast to compute the data. Some researchers want supercomputers to be equipped with their own AI capabilities.
We live in an age where researchers with their own supercomputers can come up with the fastest results.
We are currently working with researchers at the University of Tokyo campus and at RIKEN, which collaborates with the University of Tokyo, to develop specialized chips that meet everyone's needs.
Eventually, students with research experience using specialized chips will enter the workforce. They will bring research methods using specialized chips to industry. In this way, I hope that the need for agile, dedicated chips will gradually permeate the industry.