On May 23, TSMC held the “2024 Technology Forum Taiwan Station” event to share TSMC's latest technological progress and production capacity layout, and also confirmedTSMC's Nanjing plant recently obtained an “indefinite exemption” license from the US Department of Commerce.
By 2030, 3D packaging will enable integration of 1 trillion transistors on a single chip
At the technical forum that day, TSMC Asia Pacific Business Director Daniel Wan gave an opening speech. He said that looking forward to future AI innovation, high performance, 3D chip stacking, and packaging technology are becoming increasingly important. TSMC expects to achieve more than 200 billion transistors on a single chip in the next few years, and more than 1 trillion transistors through 3D packaging, which will be an exciting breakthrough in semiconductor technology.
Wan Ruiyang said that we are currently witnessing the advent of a new era of AI. AI is changing the world by analyzing large amounts of data, making predictions and automating prediction decisions. It has already tracked changes in icebergs in satellite images at a speed 10,000 times faster than humans, and has identified genes that are difficult to detect through millions of proteins.It is expected that there will be 100,000 generative AI humanoid robots by 2030, and global shipments of generative AI mobile phones are expected to reach 240 million units by the end of this year.
With the help of powerful cloud AI processors and the development of terminal-generated AI,Disruptive innovation of AI has already ushered in the fourth industrial revolution.The first was driven by steam engine mechanization, the second was driven by electrification and large-scale production, and the third was driven by semiconductor technology, which brought computers and automation. With the advent of the new era of AI, the most advanced AI uses the world's most advanced process technology, and the large language models used for AI training by TSMC's 4-7nm advanced process have increased complexity and capabilities. Therefore, in order to maintain continuous large-scale training, stronger computing power and better energy efficiency are required.
Wan Ruiyang pointed out that in addition to the demand for high-performance computing platforms brought about by generative AI, computing power demand is also seen in automotive electronics. High-performance computing is crucial to promoting L4 and L5 autonomous driving solutions. This will require 5nm or even 3nm advanced logic technology. TSMC will also continue to challenge the limits of process miniaturization.
Wan Ruiyang said that looking forward to the future, in order to meet the huge demand for high-performance computing for AI innovation, 3D chip stacking and advanced packaging technology will become increasingly important. TSMC is a global leader in advanced manufacturing.We expect to integrate more than 200 billion transistors on a single chip in the next few years, and reach more than one trillion transistors through 3D packaging.This would be an exciting breakthrough in semiconductor technology.
At the same time, Wan Ruiyang pointed out that TSMC has also promoted countless innovations. Through close cooperation, it has created a win-win strategic alliance, unleashed more powerful AI with leading semiconductor technology, achieved seemingly impossible innovations, and made the world a better place.
AI chip demand is expected to grow 2.5 times this year
In his keynote speech, TSMC's senior vice president of Europe and Asia business and deputy co-chief operating officer Hou Yongqing said that the demand for AI is strong and the demand for AI chips is expected to grow 2.5 times annually. He hopes to work with partners to face the new era of AI full of golden opportunities.
Hou Yongqing said that the industry is gradually recovering and the most difficult part has passed. The smartphone and PC markets are slowly recovering and are expected to grow by 1-3% annually. The demand for AI/HPC data centers remains strong.The demand for AI accelerators is expected to grow 2.5 times annually;Demand in the smart car chip market remains weak and will decline by 1% to 3% year-on-year this year; the Internet of Things market is expected to grow by 7-9% this year, but this is relatively weak compared to the previous 20%.
TSMC mentioned in its April earnings conference that server AI processors will more than double their revenue this year, accounting for a low-teens percentage of TSMC's total revenue in 2024. Server AI processors are expected to grow at a 50% compound annual growth rate over the next five years, and will account for more than 20% of TSMC's revenue by 2028.
Hou Yongqing said TSMC expects the global semiconductor market, excluding memory chips, to grow 10% year-on-year this year.The global foundry market is expected to grow 15-20% year-on-year (not including Intel IFS)This statement is consistent with TSMC's previous earnings conference.
In addition, Hou Yongqing expressed that with the strong growth of the AI/HPC data center market, TSMC continues to expand its ecosystem partners and has added memory, substrate, packaging and testing partners to the system and cooperate with partners to provide a complete integrated solution from chips to packaging.
During the forum, Hou Yongqing also used ChatGPT3.5 to demonstrate how TSMC helps customers. He also said humorously, “It seems that the relevant summaries are similar to the speech he just said. I don't know who copied whom.”
Hou Yongqing emphasized that TSMC will not compete with its customers, trust and partnership can create more business opportunities, and today is a new era of AI full of golden opportunities.
Hou Yongqing also cited data pointing out that the output value of semiconductor wafer foundry is expected to reach US$150 billion in 2024, which is expected to support a global economy of US$110 trillion. It is estimated that the output value of wafer foundry will reach US$250 billion in 2030 and support a global economy of US$150 trillion.
TSMC's 3nm production capacity will more than triple this year
TSMC senior plant director Huang Yuanguo pointed out that benefiting from the demand for HPC, AI and smartphones,This year, TSMC's 3nm production capacity will increase by more than three times compared to last year, but this is actually not enough.Therefore, TSMC is still working hard to meet customer needs, and it also faces challenges in rapidly increasing advanced process production capacity.
Huang Yuanguo pointed out that from 2020 to 2024, TSMC's advanced process revenue below 7nm will have an annual compound growth rate (CAGR) of more than 25%. Currently, TSMC's N3 yield is almost the same as N4. This has also prompted leading customers to prefer N3 processes for high-end chips. In addition, the special process will have a CAGR of 10% from 2020 to 2024, among which the annual CAGR of automotive chip shipments is as high as nearly 50%.
Specifically, TSMC's initial first-generation 3nm process (N3, also known as N3B) had a short life cycle, and Apple was the only major customer. The second-generation 3nm process (N3E) is a relaxed version of N3B, which eliminates some EUV layers and sacrifices some transistor density, but helps reduce production costs and expand process windows and yields.
N3E has started mass production as planned in the fourth quarter of last year. Currently, the D0 defect density of N3E is comparable to that of N5. It has been adopted by several major customers, and the related customers' products have achieved excellent yield performance.
As for N3P, which will be mass-produced in the second half of this year, this is the most advanced process in TSMC's 3nm family. Compared with the previous generation N3E process,At the same leakage rate, the N3P process can improve performance by 4%, or reduce power consumption by 9% at the same main frequency.The overall transistor density has increased by 4%.
TSMC said that N3P has completed certification and its yield performance is close to that of N3E. In addition, since N3P is an optical miniaturized version of N3E, IP modules, process rules, electronic design automation (EDA) tools and design methods are compatible with the former. Therefore, TSMC expects that most of the newly launched chip Tape-Out will use N3P instead of N3E or N3, because N3P has lower cost than N3 and higher performance efficiency than N3E. Based on the advantages of N3P, it is expected that the new products of most customers who originally used N3/N3E will be transferred to N3P process.
In 2025, TSMC's fourth-generation 3nm process (N3X) and 2n (N2) will both be mass-produced. Compared with N3P, N3X node chips can reduce the operating voltage (Vdd) from 1.0V to 0.9V, reduce power consumption by 7% at the same clock speed, or increase computing performance by 5% at the same chip area, and increase transistor density by about 10% at the same clock speed. Compared with previous generation products, N3X has the advantage of a maximum voltage of 1.2V, which is very important for ultra-high-performance applications such as desktop PCs or data center GPUs.
The N2 process node is TSMC's first GAA nanosheet transistor node, which will significantly enhance computing performance, power consumption, and chip area (PPA). Compared with N3E, the transistor density of the N2 node has increased by 15%, and the power consumption of the chip can be reduced by 25%~30% with the same number of transistors and main frequency; or the performance can be improved by 10%~15% with the same number of transistors and power consumption.
Regarding the progress of TSMC's latest N2 process, Kevin Zhang, senior vice president of business development at TSMC, pointed out that the 2nm process is currently progressing smoothly and using nanosheet technology. The current nanosheet conversion performance has reached the target of 90% and the conversion yield is over 80%. The progress is smooth and it is expected to achieve mass production of the technology in 2025.
However, it should be pointed out that although the performance, power consumption and transistor density of the N2 node are better than those of the N3 process overall, the high-voltage version of N3X is still likely to challenge N2 in performance, and since N3X is still a FinFET transistor, it will have more advantages in cost and cost-effectiveness.
In 2026, TSMC will mass produce N2P and A16 (1.6nm). Compared with the first-generation N2, N2P can reduce power consumption by 5%~10% with the same main frequency and number of transistors, and improve performance by 5%~10% with the same power consumption and number of transistors.
TSMC's A16 process will bring a new super rail (back power supply) technology. Compared with the N2P process, the A16 chip density is increased by up to 7-10%. At the same Vdd (operating voltage)/power consumption, the performance can be improved by 8-10%; at the same performance, the power consumption is reduced by 15-20%, which can support higher performance data center products.
Zhang Xiaoqiang said that A16 is the beginning of the angstrom era.TSMC's innovative backside power supply technology moves the power supply circuit to the backside.It does not affect transistor density and design, and makes the design more flexible and efficient.
Regarding the progress after the A16 process, Zhang Xiaoqiang pointed out that future transistors will have continuous innovation and evolution in research and development. The next generation architecture is likely to be from Nanosheet to CFET architecture, and material innovation is becoming more and more important. The CFET architecture is no longer just talk. The R&D team successfully verified it last year and published it by the TSMC R&D team at the 2023 IEDM. This is an important milestone in laying the foundation for the future development of semiconductors.
TSMC's capacity layout: 7 plants under construction this year
Due to strong demand for AI and HPC, TSMC is also actively expanding its advanced process and advanced packaging production capacity.
According to reports, TSMC built five new factories from 2022 to 2023, and has seven factories under construction this year, three of which are advanced process wafer fabs, two are packaging plants, and two are overseas wafer fabs.
Specifically, in terms of wafer foundry, on the island of Taiwan, China, the newly built Hsinchu Fab 20 and Kaohsiung Fab 22 are both TSMC's 2nm wafer fabs. They are progressing smoothly, have already started installation, and are expected to go into mass production in 2025.
In terms of overseas wafer fabs, TSMC plans to invest $65 billion to build three cutting-edge process wafer fabs in Arizona, the United States. The first wafer fab has already started installation and is expected to mass produce 4nm next year; the second wafer fab, which will start construction at the end of 2022, is expected to mass produce 3nm in 2028; the third wafer fab is still under planning and is expected to enter mass production before 2030.
In Kumamoto, Japan, TSMC plans to build two wafer fabs. The first wafer fab in Kumamoto will start construction in April 2022 and is expected to mass produce 22/28nm and 12/16nm processes in the fourth quarter of this year; the second Kumamoto fab has not yet started construction and is expected to mass produce 6/7nm processes in 2027.
In Dresden, Germany, TSMC will build a 16nm wafer fab, which is expected to start construction in the fourth quarter of this year and go into mass production in 2027, mainly to meet the needs of European customers;
In mainland China, TSMC Nanjing Fab 16 has expanded its production capacity from 20,000 wafers of 16nm/14nm advanced process technology toExpand production by 20,000 pieces of 28nm process.The expansion project started mass production in the second half of 2022 and will achieve full production in 2023.
As for the advanced packaging plants currently under construction, TSMC's AP5 plant in Taichung is responsible for mass production of CoWoS and is expected to start mass production this year. The AP7 plant in Chiayi will be built this year.Mass production in 2026, responsible for mass production of SoIC and CoWoS.
According to TSMC's planned SoIC and CoWoS production capacities, the compound annual growth rates (CAGRs) from 2022 to 2026 will exceed 100% and 60% respectively.
TSMC's Nanjing plant granted indefinite exemption
On October 7, 2022, the United States introduced a new semiconductor export control policy to China, restricting the ability of wafer manufacturers located in mainland China to obtain advanced semiconductor manufacturing equipment unless they obtain permission from the U.S. Department of Commerce. This includes foreign-funded companies such as Samsung and SK Hynix, as well as Taiwanese companies such as TSMC, which have wafer factories in mainland China.
Although Samsung Electronics, SK Hynix, and TSMC all received a one-year exemption from the U.S. Department of Commerce a few days later, they can obtain the supply of U.S. semiconductor equipment within the next year without any additional procedures, which also means that the production of their factories in mainland China will not be affected by the ban for the time being. However, whether the exemption can be successfully extended in the future has become the primary issue for their future development in China.
On October 9, 2023, the South Korean presidential office announced that the U.S. Department of Commerce has agreed to provide an “indefinite exemption” to Samsung Electronics and SK Hynix's wafer factories in China, which means that U.S. suppliers can supply semiconductor equipment to Samsung and SK's wafer factories in China without any license.
Then on October 13, 2023, the Ministry of Economic Affairs of Taiwan, China, confirmed that TSMC's wafer factory in mainland China had obtained a one-year exemption from the US Department of Commerce. TSMC's Nanjing plant, which has advanced process production capacity, did not receive the same “indefinite exemption” as Samsung and SK Hynix, which also makes its future development still uncertain. According to Taiwanese media reports, this one-year exemption will expire on May 31 this year.
On May 23, 2024, TSMC announced that the U.S. Department of Commerce had recently issued a “certified end-user” authorization to TSMC (Nanjing) Co., Ltd., replacing the temporary written authorization issued by the Department of Commerce since October 2022, confirming that items and services covered by U.S. export control regulations can be provided to TSMC Nanjing Factory on a long-term basis. Suppliers do not need to obtain individual licenses to supply goods, and the Nanjing Factory is expected to maintain the status quo.
However, TSMC also pointed out thatThis VEU authorization does not add any new permissions, so it can only maintain the status quo of TSMC's Nanjing plant.