Will Chiplet become a breakthrough for China's semiconductor industry?

On May 12, the third theme event of "Zhangjiang Hi-Tech 895 Entrepreneurship Camp & Aijiwei·Future Car Special" co-organized by Zhangjiang Hi-Tech 895 Entrepreneurship Camp and Aijiwei was held in the form of online live broadcast. According to Aijiwei statistics, the total number of online viewers of this event was 33,200, including 25,276 on Aijiwei official website & APP, 6,450 on Jiwei official Weibo, 157 on Bilibili, and 1,307 on video accounts.

Guests invited to this event include Yi Hui, deputy general manager of Zhangjiang Hi-Tech Venture Capital Management Department, Yang Yuxin, chief marketing officer of Black Sesame Intelligence, Liu Yibo, head of Yixing Intelligent Operations and Supply Chain and semiconductor packaging technology expert, Chen Ying'an, vice president of R&D of Xinqiyuan, Li Jue, VP of market strategy of Xinlai Technology, Zhu Jundong, vice president of products and solutions of SGD Moore, and Meng Fansheng, chief marketing officer (CMO) of Xinyuan Semiconductor. The event was hosted by Chen Yuenan, senior analyst of Jiwei Consulting.

During this event, the guests shared their insights on the industrialization of the chiplet design concept and engineering method, such as key technical bottlenecks and major application areas, and provided solutions and coping strategies for the problems currently faced by industry insiders in related industries.

Yi Hui said in his speech that Chiplet is increasingly seen as an important way to continue Moore's Law. In March this year, Intel and other ten industry giants jointly launched the UCIe Alliance to jointly promote the standardization of chiplet interface specifications. Chiplet is an inevitable trend and can even be expected to become one of the breakthroughs for China's semiconductor industry. This event invited relevant companies including 895 Entrepreneurship Camp members Yixing Intelligence and Black Sesame Intelligence to gather together to discuss the development of China's semiconductor industry in this technology field, as well as some challenges we face in terms of technology foundation and industrial ecology.

Current Status of Chiplet Technology Development

Regarding the current development of chiplets, Chen Yuenan, senior analyst at Jiwei Consulting, said that the advancement of chiplet technology requires integration with the entire industry chain, including chip foundries, packaging and testing plants, packaging and testing materials and substrates, and equipment manufacturers. The road to building its ecosystem is quite long.

Liu Yibo pointed out that from the existing practices that embody the Chiplet concept, such as AMD's EPYC server chip, three major benefits can be obtained. First, separable functions such as I/O and memory controllers can save development resources and shorten development and market launch time by using fully verified mature solutions; second, through flexible configuration of different functions, product SKUs can be enriched and market segments can be flexibly covered; third, the differentiated advantages of different Fabs in different processes and different functions can be effectively utilized.

Liu Yibo concluded that the chiplet concept was mainly used in high-performance computing-related scenarios in the past. Large-scale applications still face bottlenecks such as high-quality hard IP, supporting EDA software, die-to-die interconnect technology, packaging materials and technology, and testing technology. Industry standards have yet to be unified and improved. Currently, only a few giants with the ability to mobilize the industrial chain are practicing it. In the future, it is necessary to establish unified industry standards and be able to form actual interest-driven measures to allow more manufacturers to participate in ecological construction. Only in this way can a virtuous circle be formed, the needs of heterogeneous integration can be met at low cost, and diversified applications can be further expanded.

Chen Ying'an reviewed his experience in participating in the development of the MoChi interconnect architecture while at Marvell, pointing out that manufacturers are gradually moving from private standards and private protocols to open standards and protocols in die-to-die interconnect technology, such as the birth of the UCIe Alliance. As industry open standards are established, more and more manufacturers will be able to participate in the ecosystem, and the concept of future system manufacturers will be the integration of different bare chips.

Li Jue said that the company has been paying attention to the chiplet field. Chiplet may change IP licensing from soft core to hard core and then to die, and Xinlai Technology is also making arrangements. As a RISC-V architecture CPU IP manufacturer, the company can get in touch with user selection needs at a very early stage. Li Jue pointed out that many manufacturers in the automotive field are currently interested in chiplets, but the cost of integration, especially the integration of RF chiplets, is currently a major bottleneck for application, and innovation in business models is needed.

Future application potential of chiplets

Regarding the possible application areas of chiplet technology in the future, Meng Fansheng predicted that chiplets are expected to be widely used in typical scenarios such as autonomous driving, data centers, and artificial intelligence. In the future, chiplet technology will also realize the interconnection of dies with different computing architectures, different process platforms, and even different process media, and adapt to the overall scenario.

Meng Fansheng believes that in the early development of chiplet technology, manufacturers in the manufacturing process have greater say. The UCIe protocol initiated by Intel can fully support the heterogeneous integration of different storage devices such as DRAM and ReRAM to form cache consistency, and there is a great opportunity to achieve more innovations on the application side.

Yang Yuxin also shared his insights on this topic. As a company deeply engaged in the field of autonomous driving, Heizhima Intelligent predicts that autonomous driving chips with computing power as the core may evolve into a single "central brain" for cars. The chip functions will be extremely complex, integrating core functions from computing to control and transmission. Chiplets will create such possibilities. In addition, there are more and more domestic semiconductor startups that make "big chips" such as CPUs and GPUs. With more functional integration requirements and higher performance requirements, the challenges faced by design are also increasing. Chiplet can achieve the separation of different functional modules and evolve in stages according to their respective optimal iteration rhythms, effectively reducing the difficulty of research and development.

Yang Yuxin pointed out that the future development of chiplets must not only solve technical problems, but also explore innovations in business models and industrial chain division of labor. The market capacity of autonomous driving chips is expected to be large, with annual chip shipments of hundreds of millions, which is enough to support a closed loop business model. However, for some categories with smaller market capacities, whether they can support chiplet technology is still a challenge.

As a manufacturer in the field of autonomous driving, Liu Yi-bo of Yixing Intelligence analyzed that autonomous driving chip companies have two development paths. One is to pursue extreme technology and use various means to explore possible computing boundaries. Companies in this field have a high willingness to try new technologies, and of course they also face the challenge of head-on competition with giants such as NVIDIA.

Yixing Intelligent focuses more on commercial implementation, such as the assisted driving function provided by Tesla, which greatly improves consumers' driving experience and maximizes its product sales. In such a track, vehicle manufacturers have a great demand for assisted driving solutions with controllable costs and L2 or L2+ performance.

Liu Yibo predicts that in the next three to five years, relevant solutions will be popularized in the price range of 150,000 to 250,000 yuan, which is the largest vehicle price range. Vehicles in this price range will be very sensitive to costs. The "high-end" domain controller solutions from NVIDIA and other companies cost tens of thousands of yuan, which is difficult for vehicle manufacturers to digest. If the Chiplet method can implement relevant solutions at a lower cost, it will be useful.

However, Liu Yibo also mentioned that the inherent properties of automotive electronics are different from those of consumer electronics. Safety and reliability are still the top priority, and the process of introducing new technologies and new architectures will not be particularly radical. It may take Chiplet to mature in other application fields, improve the ecosystem, and verify the technology before it can gradually penetrate automotive electronics.

Future Application Challenges of Chiplet

Zhu Jundong said that GE Healthcare is a company that focuses on the research and development of chiplet-related technologies. The challenges of chiplet can be summarized into three points: how to calculate quickly, how to calculate well, and how to calculate cheaply?

To calculate faster, on the one hand, we need to solve the problem of how to improve the integration of computing power and provide as much computing power as possible in the smallest possible space. Due to the limited area of ​​a single chip, multi-chip arrays and chiplets were born. On the other hand, the computing power cores that require advanced processes in the "big chip" do not actually occupy much area, usually less than 40%. 3D Chiplet can arrange other modules such as cache and core computing modules in layers to maximize the area of ​​the latter.