Domestic FPGAs enter the high-end market.

Compared with CPUs and GPUs, which have shipment volumes in the billions and market sizes in the hundreds of billions of dollars, FPGAs seem somewhat "niche" - their global market size is only in the tens of billions of dollars.

However, in the journey of domestic chip autonomy, this seemingly unremarkable chip carries significant weight and often becomes the focus of public discussion.

 01 FPGA, a Battleground

This year marks the 40th anniversary of the birth of the first commercial field FPGA. At that time, it introduced the concept of reprogrammable hardware for the first time. By creating "hardware as flexible as software," the reprogrammable logic of FPGAs completely changed the landscape of semiconductor design.

FPGAs mainly have three major characteristics: high programmable flexibility, short development cycle, and parallel computing.

First of all, different from the fully customized circuits of ASICs, FPGAs belong to semi - customized circuits. Theoretically, if the scale of the gate circuits provided by an FPGA is large enough, any logic functions of ASICs and DSPs can be realized through programming. In addition, programming can be repeated, unlike ASICs, which cannot be modified after design. Therefore, FPGAs also have relatively high flexibility.

Secondly, the manufacturing process of ASICs includes multiple steps such as logic implementation, wiring processing, and tape - out. In contrast, FPGAs do not require wiring, masking, and customized tape - out, which simplifies the chip development process. The average design cycle of traditional ASICs and SoCs is 14 to 24 months, and the development time using FPGAs can be reduced by an average of 55%. Xilinx, the world's largest FPGA manufacturer, believes that being faster is more important than being cheaper. If a product is launched six months late, it will lose 33% of its profit within five years, and for every four - week delay, it means a 14% loss of market share.

Finally, FPGAs belong to parallel computing and can execute multiple instruction algorithms at once. In contrast, traditional ASICs, DSPs, and CPUs are all serial computing, which can only process one instruction set at a time. Therefore, in some special tasks, the parallel computing efficiency of FPGAs is higher than that of serial computing.

The popularity of FPGAs is not a single - point breakthrough but the result of the combined effects of three technological revolutions: mobile communications, artificial intelligence, and the industrial Internet.

The commercialization of 5G has become the first explosion point. The number of channels that a 5G network needs to process is more than 10 times that of a 4G network, and it needs to support flexible bandwidth adjustment and multi - standard compatibility. Using FPGAs to implement base - station baseband processing can significantly shorten the R & D cycle and reduce the early deployment cost.

The rise of AI edge computing has taken FPGAs to the next level. Cloud - based training relies on the large - scale parallel computing of GPUs, but real - time inference at the edge end values low latency and low power consumption more. For example, FPGAs are particularly suitable for accelerating AI inference in IoT devices such as smart cameras and sensors.

The digital transformation of the industrial sector has opened up long - term growth space. Devices such as industrial robots and smart sensors need to process a large amount of non - standard data and have strict requirements for reliability. The radiation resistance and wide - temperature characteristics of FPGAs make them an ideal choice for industrial control.

 02 High - end FPGAs, Crowded with Leading Companies

In the market landscape of FPGAs, the dominant position of American companies has long been ingrained in the industry. Xilinx and Altera have built a high wall of duopoly, and Lattice and Microchip follow closely to divide the market. Together, these four companies account for more than 90% of the market share.

In the domestic market, three listed companies, Unisoc Microelectronics, Fudan Microelectronics, and Anlogic, form the first echelon and are striving to catch up in the race.

From communication equipment to high - definition video processing, from data centers to industrial control, high - performance FPGAs are the key nodes that cannot be bypassed. They are not only "universal controllers" that can be flexibly reconfigured but also the "throats" that control system efficiency, data throughput, and security isolation.

However, in the specific classification of FPGA products, the mid - to high - end market is still almost the exclusive territory of international giants, and domestic solutions still lack the "hardcore strength" to compete head - on.

So, what are the evaluation criteria for mid - to high - end FPGAs? How is the progress of international manufacturers?

Process technology is the standard for distinguishing different generations of FPGAs and is also the first indicator to consider when evaluating FPGAs. As a type of digital chip, FPGAs themselves follow Moore's Law, and a new generation of products is usually launched every 2 - 3 years on average. Using a more advanced process can reduce power consumption, shrink the chip size, and lower the unit cost, making the performance of the new - generation FPGAs generally better than that of the previous generation. Therefore, the process technology should be the first consideration when evaluating FPGAs.

Currently, Xilinx's most advanced product, "Versal," uses TSMC's 7nm FinFET process.

Altera's Agilex 3 series of FPGAs is based on Intel's 7nm process.

The number of logic units represents the basic capacity of an FPGA and is currently the unified indicator for evaluating the basic capacity of FPGAs. The smallest functional unit of an FPGA is called the basic logic unit, which includes a LUT and a register. The basis for an FPGA to achieve programmability is the LUT, which can implement combinational circuits by itself and can complete sequential circuits when combined with a register. That is, a logic unit has the ability to perform all digital circuit functions. Therefore, the more logic units an FPGA has, the larger its capacity and the larger and more complex the circuits it can construct. Large - capacity FPGAs directly reflect a manufacturer's technical capabilities. Since the number of logic units in large - scale FPGAs is generally above 1 million, when the number of logic units exceeds 1 million, the FPGA architecture, including the LUT, CLB, and interconnections, needs to be changed; otherwise, power consumption and latency will soar. In addition, it also requires the iteration of the supporting EDA tool design process and placement and routing algorithms. Currently, among the world's top five FPGA manufacturers, only Xilinx and Altera have the ability to continuously provide large - capacity FPGA product lines.

Currently, the FPGA with the largest capacity in the world is the VP1902 (Versal Premium) launched by AMD in June 2023, with a staggering 18,507k (18 million) logic units.

In April this year, Altera announced the official mass production and shipment of the Agilex 7 M series of FPGAs. As the industry's first high - end FPGA that integrates high - bandwidth memory (HBM2E) and supports DDR5/LPDDR5 technology, this product, with a total memory bandwidth of 1TB/s and a core configuration of 3.8 million logic units, provides disruptive computing power support for scenarios such as AI inference, 5G communication, and 8K video processing.

From the perspective of process technology, the most advanced process of domestic FPGAs is currently 14/16nm.

In terms of capacity, Chinese manufacturers have relatively mature technologies for low - capacity FPGAs. Low - capacity FPGAs refer to FPGA products with less than 100k logic units. In fact, most low - capacity FPGAs have less than 10k logic units and are mainly used in the consumer electronics field, such as LED displays and bridging, as well as in some scenarios for reservation or function expansion. Currently, the processes of domestic low - capacity FPGAs are mainly concentrated at the 55nm, 40nm, and 28nm nodes, and most of them were launched in 2019 or earlier, often being the first - generation products of domestic FPGA manufacturers.

For example, the Logos series of Pango, launched in 2017, is a 40nm low - power, low - cost FPGA with 12 - 102k logic units; Anlogic's 55nm Eagle4, launched in 2016, has 20k logic units and is mainly used in servo control and high - speed image interface conversion; Gowin Semiconductor's 55nm FPGA LittleBee, launched in 2016, is the company's first - generation product, with 1 - 8k logic units.

In the mid - to low - capacity market at 28nm, Chinese FPGA manufacturers already have mature products. Mid - capacity FPGAs mainly refer to those with 100k - 500k logic units, and their applications are mainly concentrated in the air - interface side of wireless communication, industry, automotive, and A & D fields. The mid - capacity market does not pursue the highest performance; performance and power consumption are equally important, and there are also certain requirements for cost. For example, Pango, Anlogic, and IntelliFPGA all launched 28nm FPGA products in 2020, mainly targeting Xilinx's 7 series products.

In addition, some manufacturers have launched 22nm FPGAs to replace some mid - to low - capacity 28nm FPGAs. For example, Gowin's Aurora V, launched in September 2022, is its 22nm FPGA product, with 138k logic units.

Currently, high - capacity FPGAs are the bottleneck for domestic production.

 03 High - end FPGAs, an Important Opportunity

The global high - end FPGA market is mainly dominated by Xilinx and Altera. Representative products include Xilinx's 7nm ACAP Versal, 16nm Virtex Ultrascale+, as well as Altera's 10nm (Intel 7) Stratix 10 and Agilex. These products are expensive and represent the highest level of current FPGA performance, density, and integration.

High - end FPGAs are not only the core carriers of profit and competitiveness but also a necessity for technological iteration.

On the one hand, in the low - end market, due to the low entry barriers, intense competition among enterprises has compressed the profit margins, and it is difficult to form a stable advantage. In contrast, the high - end market, with its high added value, has become the "main battlefield" for manufacturers. Taking Altera as an example, the revenue of its high - end Stratix series accounts for more than half of the total. This structural profit distribution determines that if an enterprise wants to dominate the industry, it must make breakthroughs in the high - end field. At the same time, the technical barriers of high - end products can effectively block new entrants and provide support for enterprises to build long - term competitive advantages.

On the other hand, high - end FPGAs are a "must - have" for the upgrading of emerging industries. With the rapid development of fields such as 5G communication, artificial intelligence, and autonomous driving, higher requirements have been put forward for the performance, integration, and flexibility of chips. High - end FPGAs use 20nm or more advanced processes, have a logic unit count exceeding one million, and integrate heterogeneous components such as CPUs and high - speed interfaces, which can meet the requirements of high - frequency data processing and collaborative computing in complex scenarios, which are difficult for low - end products to achieve.

Domestic FPGAs still have a gap compared with the high - end market, but domestic related enterprises have initially entered this field.

In terms of process technology, FPGAs are currently accelerating the migration to the 16nm and more advanced nodes, and the competitive advantage brought by the technological gap is becoming more and more obvious. Domestic FPGA companies have achieved relevant results at 16nm and below. Efinix launched its 16nm FPGA Titanium series in July 2020, with a maximum of 176k logic units, becoming the first domestic FPGA product in the 16nm field. According to the news on the official website of Unisoc Microelectronics, Pango has started the R & D of a 14nm high - end FPGA with hundreds of millions of gates. Fudan Microelectronics started the R & D of 14/16nm products in 2021. In 2023, its 1xnm - process FPGA with billions of gates completed small - batch trial production, entered the user - testing stage, and achieved small - scale sales.

High - capacity FPGAs with more than 500K logic units are still the bottleneck for domestic production. Fudan Microelectronics' 28nm FPGA with billions of gates is the first of its kind in China, containing approximately 700K logic units, integrating SerDes (up to 13.1Gbps), DDR4, a hard - core ARM, and an AI acceleration module. It targets Xilinx's Zynq series and is applied in fields such as communication core networks, medical equipment, and automotive electronics.

Wuxi Zhongwei Yixin specializes in the R & D of high - performance FPGAs. Its products include the YX5F200T and are applied in industrial control and aerospace fields. In 2024, its 16nm FPGA was successfully trial - produced, with a logic unit density of 500K.

The localization of FPGAs is a tough battle that needs to be advanced steadily. As an important part of the digital hardware field, its significance in the process of localization is not diminished by its smaller market size compared with CPUs and GPUs.

This article is from the WeChat official account "Semiconductor Industry Insights" (ID: ICViews), author: Feng Ning, published by 36Kr with authorization.