At the moment of lack of normalization of core, we find an interesting fact: the cost of mature process/old process is also increasing. We should know that with the increase of yield and output in each generation of manufacturing technology, we can dilute the capital investment of CapEx in the early stage by quantity, and each generation of technology should be cheaper and cheaper. The rare increase of the old process cost is unprecedented in the industry in the past. It can be seen that the core shortage in the current market has brought great changes to semiconductor manufacturing. ... Two years ago, combined with a paper from MIT, we wrote an article on the trend of semiconductor industry called "The Rise of Deep Learning, the Elegy of General Computing? 》。 Looking back at MIT's paper now, I feel like a standard. There is an important basic theory, that is, with the increasing cost of cutting-edge technology in semiconductor manufacturing, the market becomes more and more fierce. Five years ago, the American Semiconductor Industry Association also estimated that the supporting manufacturing cost of 7nm technology nodes would cost 7 billion. In a blink of an eye, we can see how much the capital expenditure of TSMC is expected to be CapEx this year, and we can see how exaggerated the cost of cutting-edge manufacturing technology is. So what will happen if the cost increases? No matter how the growth rate of an industry or an industry is limited, if the cost growth rate exceeds the expected development rate of the industry itself, it will not only lead to the price increase of products and the slowdown of technology iteration cycle, but also lead to the market participants having to eat into other competitors to survive, and the oligopoly effect will become increasingly prominent. This is the case in the semiconductor manufacturing market in the past two years. There is also a message implied here, that is, the de facto end of Moore's Law. Some people will say: This is a cliche. If we don't look at Moore's Law from a technical point of view, Moore's Law not only means that the number of transistors per unit area doubles every 12-18 months, but also that the cost of achieving the same performance will theoretically be reduced. There are a lot of people discussing whether the first half is right or not, and the second half may not be reached. Another interesting thing is that if you think it is not uncommon for the cost of cutting-edge manufacturing technology to rise, then the cost of mature technology (old technology, also called mainstream technology by some people) is also rising now. Does that sound strange? Recently, Fabricated Knowledge has published an analysis of the rising cost of semiconductor manufacturing, which is quite interesting in combination with the current trend of core shortage. At the same time, if we combine the upstream of semiconductor manufacturing and large silicon wafer manufacturers, we can actually better understand what the general trend of "transistor price increase" is and how long it will last.
Cutting edge technology and Moore's law For the argument of "Moore's Law is stagnant", there are naturally many manufacturers who oppose it. For example, ASML, such as major EDA manufacturers, such as several foundry factories; In the past two years, Synopsys has spared no effort to publicize SysMoore and promote the development of Moore's Law to the systematic level. Foundry factory focuses on more than Moore, that is, the development of advanced packaging technology. The improvement of performance and efficiency continues in the semiconductor industry, which should not be a false proposition. However, these market participants didn't say much about whether the cost of achieving the same performance is increasing. We should know that the cornerstone of the development of the whole field of information technology and even electronic technology is largely in the process of short technical iteration period and decreasing cost. To put it bluntly, this is the driving force of the third scientific and technological revolution.
Fabricated Knowledge cites a picture of Marvell's Investor Day in the previous two years. This figure shows the cost per hundred million logic gates under different process nodes. It can be clearly seen that 28nm is a critical point. 28nm is actually the critical point between the planar FET transistor and the FinFET transistor. Since then, the cost per hundred million gates of each generation of technology has been slowly rising. For contemporary manufacturing technology, we are also at a critical point: in terms of cutting-edge manufacturing technology, GAAFET transistor will soon replace FinFET-next year, Samsung GAAFET transistor will take the lead in mass production, and TSMC and Intel will follow suit. Does this mean that there will be a new wave of cost increase after the 3nm node? Because the manufacturing of GAAFET transistor requires more complicated and diversified process steps. In fact, in 2018, someone wrote an analysis paper entitled Measuring Moore's Law: Evidence from Price, Cost, and Quality Indexes. It was said that if the design and manufacturing cost of a single transistor is taken as the analysis object, there will be a period of time in history when the number will drop by 20-30% every year-that is to say, process iteration can bring higher economic benefits; However, after 28nm, this trend will no longer exist, which is also the unanimous opinion of analysts now. On the issue of cost increase, CSET in the United States also wrote a paper about the rising cost of AI chips the year before last. The starting point is quite different, and the total cost including power consumption cost and operating cost is discussed more comprehensively. Readers interested in cutting-edge manufacturing technology can read it. In fact, the cost of manufacturing equipment has increased significantly in recent years (from 2012 to 2020, the market increase of wafer manufacturing equipment is obviously greater than the total shipping capacity), and the price is still rising; Not only the price of EUV mask aligner, but also all kinds of new technologies to achieve higher density chip manufacturing (such as 3D NAND) and add more processes, all of which are raising the manufacturing cost. New technologies such as advanced packaging technology are generally increasing the manufacturing cost of a single transistor.
The cost of the old technology is actually increasing? Another interesting fact is that the cost of mature technology/old technology is also increasing this year. Is this totally counterintuitive? We know the next generation of manufacturing processes. With the increase of yield and output, the capital investment of CapEx in the early stage should be diluted by quantity, and each generation of processes should be used more and more cheaply. When it becomes an old process, there is an improved cost optimization scheme to further reduce the cost. Generally, the manufacturing factory of the old technology is depreciation and maintenance cost. However, in the past two years, the industry has faced unprecedented changes, that is, the supply of chips is in short supply, and it has reached the point of large-scale core shortage. A large part of the core shortage is the chips manufactured by the old technology, and the demand is also showing a sharp increase. Applications such as automobiles and IoT, which are seriously lacking in core, generally don't adopt cutting-edge manufacturing technology, but only use lower-cost old technology based on their quantity (the reason is already in "The Rise of Deep Learning is the Elegy of General Computing?" There is a clear explanation in the article). Generally, when the utilization rate of a fab factory reaches 80%, we can say that it is in a "full load" state; In fact, VLSI Research's data shows that the utilization rate of fab, a sub-cutting process, was nearly 100% last year. This requires the mature process to expand production: that is to say, it is now necessary to increase CapEx's actual capital expenditure, buy equipment, build factories or rebuild for the old process. The extra CapEx expenditure will inevitably lead to the price increase of these old processes in the short to medium term. Companies including Silicon Labs, Anselmei, NXP, etc. have expressed this attitude, and indicated that this trend of sub-cutting or mature processes will continue for some time due to the core shortage tide. Fabricated Knowledge commented that this was unprecedented in the industry in the past. For fab and foundry factories, they still face the difficulty of choice, that is, how much demand these mature technology chips are and whether it is really worth building a new factory (so-called greenfield investment). At least, car companies are quite determined in this respect, and they are willing to pay a higher price for mature technology, or even pay a full deposit. If the customer can place such an order to ensure that it will not be cancelled, the manufacturer will naturally invest. Of course, as time goes on, the cost of mature technology will still decrease slowly in the future. But Fabricated Knowledge thinks it may take several years, especially now that the market demand for automobiles and IoT is still so strong. Present situation of core deficiency in silicon wafer factory Lack of core is a big topic, but there are always some starting points to talk about. For example, we think that those who make silicon wafers have a great say in the problem of core shortage. SUMCO expects the compound annual growth rate (CAGR) of smartphone shipments to reach 4.9% from 2020 to 2025, increasing from 1.3 billion to 1.65 billion units. In terms of HPC (including data center CPU, AI chip, application processor of 5G mobile phone, autonomous driving CPU, etc.), it is expected that there will be 14.7% CAGR in these five years. In addition, the annual bit growth rate of DRAM is 20%, and the wafer demand is CAGR 20% in the past five years and so on. The demand for many applications seems to have no signs of slowing down at all. SUMCO and Global Wafers both express that they can't keep up with the market demand quickly because they lack key tools to increase production. Its output capacity cannot be expanded rapidly until 2023; The production capacity will continue to increase by 2025, but it is not enough to meet the demand of silicon wafers. Before 2024, the price of SUMCO silicon wafer will increase by 10% every year. In terms of mature manufacturing technology, a representative thing: on earning call, SUMCO said that it would not and could not increase the output of 200mm wafers. At present, SUMCO's supply capacity improvement is all in brownfield investment (brownfield refers to the new deployment on the old system, compared with greenfield, it is generally a brand-new investment). By 2021, the capacity increase achieved by brownfield expansion has reached its limit.
In addition, SUMCO also gave the data of 300mm silicon wafer inventory of its downstream customers. SUMCO said that the customer inventory suffered a steep drop in the past year. "We believe that the customer's inventory has dropped to 1 month's level. If the inventory is lower than the level of one month, it already constitutes a dangerous situation for customers. " If the inventory level is divided into logic and storage wafers, "you will see that the logic side has reached 0.7 months, while the storage is about 1.2 months." It can be seen how much and how long the current market situation will affect the cost of subsequent semiconductor manufacturing. Therefore, the forecast of "several years" is probably practical. Finally, it is worth mentioning that TSMC, as the leader of foundry Factory at present, has almost complete market voice in cutting-edge manufacturing technology. We have also reported many times before that TSMC expressed its confidence in improving profit margin in the future with ambition based on market demand. Part of improving the profit margin is to increase the price-whether it is to make more money or the expected market development speed mentioned at the beginning of this article can't keep up with the cost increase. In last year's article on observing the mobile phone market, we talked about the fact that TSMC adjusted the wafer price, especially the wafer price of mature process nodes, which will continue to increase this year. With TSMC's current market position, buyers will not have much choice. This also needs to be considered as a part of the cost increase in the semiconductor field.