The Future of ASIC Chips in AI Technology

ASICs are proprietary application-specific integrated circuits designed and manufactured to meet specific user requirements and electronic systems. With their advantages of high performance and low power consumption, they are widely used in various fields including electronics, computer science, healthcare, automotive industry, and artificial intelligence. The global ASIC market is expected to grow to $24.7 billion by 2025, with a compound annual growth rate (CAGR) of 41.4% in the Chinese market, indicating a significant expansion. Chinese ASIC manufacturers have shown outstanding performance in technology and market competition, especially in 7nm processes and computing power, aligning with international standards. The ASIC industry chain forms a complete ecosystem from design to application. Technological advancements, particularly the application of new materials and the development of process technologies, present new opportunities for ASIC chips, showcasing greater advantages and potential in the AI field.

ASIC Chips Support Customized High-Performance Computing

ASICs are designed based on specific user needs and electronic systems, providing optimized performance, power consumption, size, and cost compared to general-purpose chips. The design and manufacturing process of ASIC chips includes multiple precise iterative steps such as requirement analysis and architectural design, allowing them to significantly outperform traditional general-purpose chips. Among the mainstream AI chips, ASICs offer the best computational efficiency, while GPUs have strong versatility, and FPGAs depend on specific design application scenarios.

1. Overview. ASIC (Application Specific Integrated Circuit) is a proprietary application-specific chip designed and manufactured to meet specific user requirements and electronic systems. Its computing power and efficiency can be customized according to algorithm needs, typically optimizing performance, power consumption, size, and cost compared to general-purpose chips.

2. Characteristics. ASIC chips typically feature high customization, high performance, compact size, and low power consumption. In terms of performance, ASIC chips generally far exceed traditional general-purpose chips. For example, in oscilloscope design, a general-purpose chip may only achieve 1GHz bandwidth, while an oscilloscope using an ASIC chip can achieve tens of GHz bandwidth. ASIC chips are meticulously optimized during the design process, eliminating unnecessary logic units and processing components, constructed in a pure digital logic circuit form, which not only helps to reduce chip area but also improves overall efficiency. In terms of energy efficiency, the unit power consumption of ASIC chips is significantly lower than that of CPUs, GPUs, and FPGAs, making them very suitable for smart appliance applications with strict energy consumption limits.

3. Design and Manufacturing Process. The design and manufacturing process of ASICs is a precise and iterative process that involves several key steps including requirement analysis, architecture design, digital design, verification, logic synthesis, physical design, post-simulation, and then proceeding to tape-out, testing, packaging, and final testing before delivery.

4. Comparison of Mainstream AI Chips. The current mainstream AI chips are mainly divided into three categories: GPUs, FPGAs, and ASICs. GPUs are known for their high versatility, capable of handling diverse computational tasks, but they are slow and low-performing when executing AI algorithms. FPGAs are semi-custom hardware that allows dynamic programming and configuration at the hardware level, providing sufficient computing power and flexibility, but they have longer development cycles, complex algorithms, and high development difficulty. ASICs can be tailored to specific product needs but have high R&D costs and average replicability. Overall, ASICs provide the optimal computational efficiency, followed closely by GPUs, while the performance of FPGAs depends on their specific design and application scenarios.

ASIC chips are categorized into fully customized, semi-custom, and programmable types. Fully customized ASICs offer the best performance, achieving eight times the computing power of semi-custom ASICs. Semi-custom ASICs combine standard logic units with custom designs. Programmable ASICs (PLD) meet specific needs through programming, providing flexibility.

1. Fully Customized ASIC Chips. Fully customized ASIC chips are among the highest level of customization. R&D personnel design logic units targeting different functions based on various circuit structures, building analog circuits, storage units, and mechanical structures on the chip board. Fully customized ASIC chips perform excellently in terms of performance and power consumption, with average computing power output approximately eight times that of semi-custom ASIC chips. A fully customized ASIC chip using a 24nm process outperforms a semi-custom ASIC chip using a 5nm process.

2. Semi-Custom ASIC Chips. Semi-custom ASIC chips are integrated circuit designs that fall between fully customized ASICs and standard products. The logic units constituting semi-custom ASIC chips are mostly derived from standard logic unit libraries, with some custom designs based on specific needs, combining performance optimization of fully customized ASICs with the flexibility of programmable ASICs. Semi-custom ASIC chips are further divided into gate array chips and standard cell chips. In gate array chips, the predetermined transistor positions on the silicon wafer cannot be changed, and designers often adjust the interconnection structure of logic units by altering the bottom metal layer of the chip. Standard cell ASICs consist of pre-designed and verified logic units that have completed layout, allowing designers to arrange existing standard units according to algorithm requirements, providing high design flexibility and faster development cycles.

3. Programmable ASIC Chips. Programmable ASIC chips are a special type of integrated circuit that combines the high performance of ASICs with the flexibility of FPGAs (Field-Programmable Gate Arrays). These chips allow designers to program or reprogram the chip’s functionality after the silicon wafer has been manufactured, adapting to different application needs or performing functional updates. Compared to fully customized ASICs, programmable ASICs have lower development costs during the development phase because they do not require new masks for each design. However, after mass production, the unit cost may be slightly higher due to the programming and verification process.

The ASIC Chip Market Continues to Expand

The global AI chip market is expected to grow from $17.5 billion in 2020 to $72.6 billion by 2025, while the Chinese AI chip market is projected to increase from 6.4 billion yuan in 2018 to 85 billion yuan in 2021. Meanwhile, the global ASIC market is expected to reach $24.7 billion by 2025, with a compound annual growth rate (CAGR) of 41.4%, driven by the development of AI technology and the rapid expansion of applications across multiple industries.

1. Global AI Chip Market Size. According to data from the China Business Industry Research Institute, the global AI chip market size is expected to increase from approximately $17.5 billion in 2020 to $72.6 billion by 2025 (nearly 500 billion yuan). This growth is primarily driven by the rapid development of artificial intelligence technology, accelerated commercialization, and the continuously expanding demand for AI chips across various industries, including smart security, autonomous driving, smartphones, smart retail, and intelligent robotics.

2. Chinese AI Chip Market Size. As a significant player in the global AI chip market, China’s market size is also continuously expanding, growing from approximately 6.4 billion yuan in 2018 to 85 billion yuan in 2021, with an average annual compound growth rate of 67.7%. Analysts from the China Business Industry Research Institute predict that the Chinese AI chip market size will grow to 230.2 billion yuan by 2024, with GPU usage being the largest, reaching a market share of 89.0% in 2022. The market shares of NPU, ASIC, and FPGA are relatively low, at 9.6%, 1.0%, and 0.4% respectively.

3. ASIC Market Size. According to KBV Research, the global ASIC chip market size is expected to reach $24.7 billion between 2019 and 2025, with an expected compound annual growth rate (CAGR) of 8.2%. According to Frost & Sullivan’s data, the sales scale of ASIC chip products in China was 6.6 billion yuan in 2018, and the market size grew to 37.3 billion yuan in 2023, with a compound growth rate of 41.4%, far exceeding the global level. Among them, semi-custom and programmable ASIC chips are designed and manufactured faster than fully customized ASIC chips.

Marvell strengthens its competitiveness in ASIC chips for 5G, cloud data centers, enterprise, and automotive applications through the acquisition of Avera Semiconductor. Alchip, a fabless ASIC design company founded by engineers from Silicon Valley, has gained recognition for its rapid implementation of 16nm to 7nm processes. Broadcom, the second largest AI chip company globally, holds a significant position in the networking and wireless sectors. GUC achieves innovation on the latest process nodes, serving communications, computing, and consumer electronics.

1. Marvell. In May 2019, Marvell announced an agreement to acquire GlobalFoundries’ ASIC business, Avera Semiconductor, which helps chip designers develop semi-custom chips within fully customized chips. Marvell aims to disrupt the custom ASIC chip market with new 5nm products targeting 5G operators, cloud data centers, enterprises, and automotive applications.

2. Alchip. Alchip is a foreign-funded enterprise founded by engineers in Silicon Valley, emerging as an ASIC design service company in the AI boom, primarily providing fabless ASIC solutions to meet customer needs in the artificial intelligence field. Alchip has rapidly achieved success at 16nm, 12nm, and 7nm process technology nodes, boasting a reliable track record.

3. Broadcom. Broadcom is the second largest artificial intelligence chip company globally, second only to Nvidia, with ASIC sales reaching billions of dollars. Broadcom not only assists in the design of ASIC chips but also provides key intellectual property for ASIC chip production, testing, and packaging. It currently has a large business in networking equipment and wireless chips, receiving up to $1 billion annually from companies like Google for manufacturing ASIC chips that run servers.

4. GUC. GUC is an advanced customized IC service company headquartered in Taiwan that closely collaborates with TSMC and major packaging and testing companies by combining advanced technology, low power consumption, and embedded CPU design capabilities. Over the years, GUC has utilized Cadence’s full digital flow to achieve challenging ASIC designs at the latest 5nm and 3nm process nodes, providing the most suitable ASIC designs for advanced communications, computing, and consumer electronics, widely applied in smart search, recommendation systems, and big data analysis.

ASIC Chips Drive Technological Innovation Across Multiple Fields

The ASIC chip industry chain includes upstream design and material supply, midstream manufacturing and testing, and downstream terminal applications. The design stage involves companies like Zhaoxin and Alibaba, while the manufacturing segment focuses on basic algorithm chips and functional ASIC modules. The downstream applications are primarily in autonomous driving, smart appliances, etc., with the greatest demand in smart security.

1. Upstream of the Industry Chain. The ASIC chip industry chain consists of upstream algorithm design companies, IP core licensing companies, EDA tool suppliers, wafer and tape-out foundries, and suppliers of specialized materials and equipment. ASIC chips essentially serve as computation acceleration engines, whose computational architecture includes a main control system and auxiliary computation system. The architecture design is usually entrusted to experienced large companies, with overseas algorithm architecture design companies including Qualcomm, ARM, Google, Intel, Microsoft, Xilinx, IBM, etc. Domestic algorithm architecture design companies include Zhaoxin and Alibaba. Among them, Cambricon has developed relatively mature computational architecture design capabilities, and its algorithms have been widely applied in Huawei’s acceleration products.

2. Midstream of the Industry Chain. The midstream of the ASIC chip industry chain includes various ASIC chip manufacturers and testing companies. ASIC chips are sold in two models: under the basic algorithm plus chip sales model, midstream companies manufacture suitable chips based on pre-set algorithms, forming a close integration between chips, software, and algorithms. In the Chinese market, the average unit price of such products is about $2, while in overseas markets, it averages around $3. In the ASIC module product sales model, companies package algorithms and chips, adding functions such as voice recognition and image recognition to create ASIC module products, enhancing their functionality. The average price of a complete ASIC module product is about $5.

3. Downstream of the Industry Chain. The downstream of the ASIC chip industry chain consists of various consumer electronics and industrial product manufacturers. ASIC chips are applied in fields such as autonomous driving, smart factories, smart appliances, and defense military. Among them, the smart security field primarily applies image recognition ASIC chips, accounting for 30% of all downstream applications.

Due to their high performance and low power consumption characteristics, ASIC chips are widely used in various fields including electronics, computer science, healthcare, automotive industry, and artificial intelligence. They enhance performance in electronic devices, accelerate encryption and decryption in computer science, support advanced equipment in healthcare, enable intelligence in the automotive industry, and optimize edge computing in artificial intelligence. The integration and specialization of ASIC chips make them key components supporting technological innovation and industry progress.

1. In the Field of Electronics. ASIC chips are widely used in electronic products, and the success of these products largely depends on the use of ASIC chips, enabling these devices to achieve higher performance and lower power consumption. ASIC chips integrate a large amount of hardware, software, and algorithms, making them faster and more energy-efficient than general processors.

2. In the Field of Computer Science. ASIC chips are applied in computer science fields such as cryptography, blockchain, etc. ASIC chips can enhance the speed of encryption and decryption, key management, and other security algorithms through hardware, making them widely used in banking, finance, cybersecurity, digital asset management, and cryptocurrency.

3. In the Healthcare Field. In the medical field, ASIC chips can be used in biosensors, medical imaging devices, medical equipment controllers, etc. In healthcare, ASIC chips have wide applications in wearable devices, health monitors, portable ventilators, and other medical applications.

4. In the Automotive Industry. ASIC chips are also used in the automotive industry, including technologies for autonomous driving, vehicle control systems, and engine control. The low power consumption and high performance of ASIC chips make vehicles increasingly intelligent and help improve automotive performance and safety.

5. In the Field of Artificial Intelligence. At the AI inference end, small size, high performance, and low energy consumption ASIC chips are very suitable. Since edge computing on IoT terminal devices needs to be performed locally, cloud computing is not suitable in this regard. Therefore, ASIC chips have become the ideal choice for edge computing on IoT terminals.

The Rapid Rise of Domestic ASICs

ASIC technology has transitioned from simple customized hardware to highly integrated system-on-chip (SoC), achieving complex single-chip integration. With technological advancements, ASIC manufacturing has evolved from micron to nanometer scale, realizing higher density circuit integration and power consumption reduction. However, ASIC design faces high development costs, complex design verification processes, and technological patent barriers.

1. Transition to System-on-Chip. The development of ASIC technology marks a leap from simple implementations of customized hardware to highly integrated system-on-chip solutions. Modern ASIC designs typically integrate components such as 32-bit microprocessors, memory units, and network interfaces, forming a so-called system-on-chip (SoC), which integrates complex functions on a single chip, greatly enhancing performance and efficiency. With continuous breakthroughs in semiconductor process technology, ASIC manufacturing processes have evolved from micron to nanometer scales, and are even moving towards finer dimensions. This advancement allows ASICs to integrate denser circuit networks in more compact spaces while achieving power consumption reduction and effective cost control. ASICs have further segmented into specialized chips such as DPU, NPU, and TPU tailored for different user needs, showcasing the strong potential of ASIC technology to meet diverse computing requirements.

2. Technological Barriers. ASIC design relies on deep expertise and technical know-how; the design and verification processes are not only complex and cumbersome but also require significant time and resource investment. Additionally, the initial development of ASICs involves expensive mask and tape-out costs, which are typically only bearable by companies with considerable scale and financial strength. The application of patented technologies is widespread in ASIC design and manufacturing processes, and these key patents are often concentrated in the hands of a few industry giants, creating high market entry barriers for new entrants.

ASIC chip technology achieves performance enhancement and cost reduction through advanced processes and new materials such as extreme ultraviolet lithography and carbon nanotubes. In market competition, domestic manufacturers such as HiSilicon and Suiyuan Technology are keeping pace with international brands, demonstrating the rapid progress of domestic ASICs and their potential to become industry leaders in the future.

1. Process Technology Innovation. Technological development will drive ASIC chip manufacturing. With the ongoing advancement of Moore’s Law, the manufacturing of ASIC chips will shift to more advanced process technologies, such as extreme ultraviolet lithography and electron beam lithography, to achieve finer circuit designs and higher integration levels. With the development of new material technologies, new materials such as carbon nanotubes and graphene will be applied in the manufacturing of ASIC chips to improve performance and reduce costs.

2. Broad Market Prospects. Unlike CPUs, GPUs, and FPGAs, the current global ASIC market has not formed obvious leading manufacturers, and domestic manufacturers are developing rapidly. Product comparisons reveal that domestic manufacturers are currently adopting 7nm process technologies, similar to foreign ASIC manufacturers. In terms of computing power, HiSilicon’s Ascend 910 surpasses Google’s latest TPUv4 in BF16 floating-point computing power and INT8 fixed-point computing power, while products from Suiyuan Technology and Cambricon also match Google’s overall performance. In the future, domestic manufacturers are expected to maintain technological advantages in the ASIC field and become leading manufacturers.

* * *

About Jiuqian Capital, established in 2009, serves enterprises and leading investment institutions focusing on emerging fields; we believe that science and technology can change professional services; we hope to bring more rationality to the market and less ambiguity; we believe that unique research and analysis is the only source of our honor.