news

한어Русский языкEnglishFrançaisIndonesianSanskrit日本語DeutschPortuguêsΕλληνικάespañolItalianoSuomalainenLatina

Specializing in edge-side smart chips

Text｜"Chinese Entrepreneurs" reporter Tao Tao

Editor: Li Wei

Image source: Interviewee

Qiu Xiaoxin wakes up at 5 or 6 o'clock every day to process work information and work in multiple threads until 11 o'clock at night before going to bed. She is 56 years old this year and has maintained her vigorous energy for 5 years since she started her business.Employees describe her as "decisive and extremely emotionally stable."

Qiu Xiaoxin is the helmsman of Aixin Semiconductor Co., Ltd. (hereinafter referred to as "Aixin Semiconductor") and a rare woman in China's semiconductor industry.

In the 1990s, after graduating with a master's degree from Tsinghua University, Qiu Xiaoxin went to the United States to study and obtained a doctorate from the University of Southern California. He then served as vice president of Broadcom and chief technology officer of Unigroup Spreadtrum.

In 2019, after leaving a large company, Qiu Xiaoxin was faced with a choice - to re-enter a large company or start his own business?

At that time,The chip industry has just experienced a process from peak to trough, and AI companies have entered a period of application bottleneck.However, Qiu Xiaoxin, who has more than 20 years of industry experience, believes that China's huge market potential and embrace of new technologies will give rise to subsequent explosions in AI applications.

With the right time, right place, and right people, Qiu Xiaoxin made a decisive decision to start his own business. In the huge AI track, Aixin Yuanzhi sought end-side and edge-side chips as a foothold.

"Over the past decade, most AI processing has taken place in the cloud. Now, with the development of AI and chip technology, the combination of cloud, edge and end constitutes the new infrastructure of artificial intelligence. The AI ​​cloud data center market has high requirements for computing power, and the customers and market are relatively concentrated, which creates a certain threshold for start-ups. In comparison, the customer concentration of edge chip customers is lower, and the application scenarios of AI are equally broad." Qiu Xiaoxin explained about entering the edge side.

In the edge market, Aixin Yuanzhi formulated a strategy to focus on the visual field at the beginning of its establishment. In Qiu Xiaoxin's view, in order for the entire industry to be AI-based, it needs data in addition to computing power.More than 80% of AI data is collected through vision, and this vertical track has the most application prospects。

Today, Aixin Yuanzhi’s two major technological moats - Aixin Tongyuan (NPU, embedded network neural processor) and Aixin Zhimo (AI-ISP), are precisely in line with edge computing applications and use visual processing, which is consistent with the company’s original strategic goals.

Unlike most semiconductor chip companies that focus on one market among mobile phones, smart driving or smart cities, Aixin Yuanzhi chooses to drive the latter two scenarios.

“The edge market is more fragmented and less concentrated than the cloud market.Extracting the commonality of the market and covering multiple scenarios with one set of technologies can effectively reduce marginal costs” Qiu Xiaoxin explained.

Looking back on the process of market expansion, Qiu Xiaoxin lamented the hardships at that time: "When the first chip was being evaluated, it was a painful process to make up for the shortcomings, just like changing the engine while flying." From the output of the second chip to winning customer orders, Aixin Yuanzhi has gone through a fierce battle in improving technology.

After five years of overcoming difficulties, Aixin Yuanzhi is now on a much smoother path. Qiu Xiaoxin said that in the second five years, Aixin Yuanzhi will continue to increase its market share and strive to achieve profitability. The ultimate goal of this company is to become a leading chip design company in China and make its products world-class.

Extracting market commonalities

When a ship sets sail, it must anchor its target. The rapid development of the edge market has provided market possibilities for Aixin Yuanzhi to choose to enter the edge market.

According to the latest report from QYResearch research team, it is expected thatThe global edge AI chip market will reach US$2.316 billion in 2024, the global edge AI chip market size will reach US$8.13 billion in 2030, with a compound annual growth rate of 16.5% in the next few years.

Cloud computing usually refers to providing centralized computing and storage resources for terminal devices through the Internet and data centers. Edge computing, on the other hand, focuses more on the side close to the data source, providing services nearby. The former is suitable for large-scale centralized data processing and analysis, while the latter is suitable for local and real-time processing of information, with the characteristics of low latency and privacy security.

"In smart driving scenarios, edge computing has great advantages. For example, if you need to step on the brakes, you cannot wait for the information to be transmitted to the cloud and wait for feedback before the system issues instructions," Qiu Xiaoxin gave an example.

In the AIoT scenario, edge computing can help the robot's visual system to do environment modeling, path planning, and action execution. "For example, if I ask the robot to get a bottle of water, it needs to understand and execute the command quickly, and the edge side and the terminal side must realize the closed loop themselves," Qiu Xiaoxin introduced.

At present, the competition for domestic automotive SoC (system-on-chip) has entered a white-hot stage. Horizon Robotics and Black Sesame are focusing on the automotive field; in the field of smart cities and AIoT, players such as Rockchip and Star Technology are competing for another piece of the pie.Aixin Yuanzhi chose to take a two-pronged approach of smart city and intelligent driving, and is committed to expanding into a broader robotics market in the future.

"The dual-wheel drive strategy is adopted because the edge and end markets are relatively fragmented, and it is difficult for each track to independently support a large-scale chip company," said Qiu Xiaoxin.

The latest survey report from ABI Research shows that the global cloud AI chip market is expected to reach US$10 billion in 2024, while the edge AI chip market will be less than a quarter of that.

The decision to enter multiple segments at the same time also comes from Qiu Xiaoxin's experience gained from years of working overseas: "Mature chip companies will make themselves platform-based companies and apply shared underlying technologies and IP to different fields. Applications on the edge are like 'elephants dancing in a box', which requires extreme performance in a limited space. Regardless of the scenario, the essence is the same. Aixin Yuanzhi's IP, chip platform, and AI computing can be applied in different scenarios."

"What we need to do is to use a set of technologies to meet the common needs of the market and improve the input-output ratio.Multiple sub-segments can also enhance a company's ability to resist risks.Qiu Xiaoxin added.

Opportunities brought by foresight

After anchoring the target, precise force becomes crucial for helmsman Qiu Xiaoxin.

In the chip field, the cycle from project establishment to output is about one and a half years, and the customer introduction cycle is from nine months to one year, totaling two to two and a half years.The long production cycle will bring higher risks to product positioning.

The company's foresight in industry development and product definition is crucial. "In product definition and design, we need to consider changes in the entire product structure or market, and how customer needs will change," said Qiu Xiaoxin. The prediction of industry trends largely determines the success or failure of a product.

Qiming Venture Partners Managing Partner Zhou Zhifeng told China Entrepreneur:When investing in Aixin Yuanzhi in 2020, the main thing I saw was their unique team.——A combination of international chip technology leaders and top AI algorithm experts. This team has the potential to achieve real disruption in the AI ​​end-side chip market, which is rich in scenarios and friendly to startups. ”

When defining the product in the early stage, Qiu Xiaoxin's team considered that the most important factors for the AI ​​of the entire industry were computing power and data, and among the massive data, visual data accounted for as much as 80%. Therefore, Aixin Yuanzhi focused its chip technology on vision. "This coincides with the previous accumulation of our R&D team," she said.

In 2020, Aixin Yuanzhi launched its first generation of chip products, which were embedded with AI-ISP technology. AI-ISP uses pixel-level AI processing technology to comprehensively improve imaging effects in various complex application scenarios, and provide high-quality images and video materials for later intelligent processing, such as making the scenery in the dark look like daytime.

In addition, Aixin Yuanzhi's other core technology, Aixin Tongyuan mixed-precision NPU, adopts a multi-threaded heterogeneous multi-core design, efficiently supports mixed-precision algorithm design, and supports Transformer network structure.It provides a good foundation for the application of large models on the edge and end sides.

At that time, between the two commonly used technologies of GPU (graphics processing unit) and NPU, Aixin Yuanzhi chose NPU after weighing the pros and cons because of NPU's relative advantages in edge and terminal computing.

"The GPU architecture is more flexible and more suitable for cloud chips. But it is too flexible on the edge, so the cost and power consumption are very high. The NPU is like a more native AI processor. It directly regards AI as a calculation method and implements these calculation methods in the most economical and reasonable way, a bit like the CPU processes addition, subtraction, multiplication and division." Qiu Xiaoxin explained.

Aixin Yuanzhi’s third-generation NPU chip, AX650N, has become the first chip platform on the domestic edge side that can support the implementation of large models."The big model became well-known after the ChatGPT series came out. But when we designed the chip at the end of 2021 and the beginning of 2022, we took Transformer support as a key consideration, so we seized the initiative," said Qiu Xiaoxin.

Changing engines on the fly

For a startup, there is a huge chasm between launching new technology, meeting basic customer needs, and gaining widespread market recognition.

Although Aixin Yuanzhi has set a reasonable goal and exerted its efforts precisely, Qiu Xiaoxin still remembersThe first chip went through a lot of challenges from tape-out to delivery.。

In July 2020, Aixin Yuanzhi's first self-developed AI chip, AX630A, was launched. At that time, Aixin Yuanzhi was full of confidence, but after customers received the chip, various problems kept popping up.

"The product has outstanding strengths, but also significant weaknesses. We relied on the team's hard work, bit by bit, to make up for the weaknesses and make the barrel leak-proof," recalled Qiu Xiaoxin.

From October 2020 to June 2021, for 8 months, Aixin Yuanzhi's account manager led a team of more than a dozen people to stay at the customer's site every day to work around the clock to solve problems.

"Every night from 10 to 12 noon, the account manager will lead everyone to hold a daily meeting, discuss each problem one by one, and come up with solutions one by one. If the customer thinks the tool chain is not easy to use, we have to refactor it.It's a bit like changing the engine of an airplane while it's flying.Qiu Xiaoxin added.

Since the production of the first chip in 2020, Aixin Yuanzhi has completed the research and development and mass production of four generations of chips. "The reason why we launch a generation of chips every year is because we need to quickly fill the product matrix and quickly optimize product performance," said Qiu Xiaoshen.

When Aixin Yuanzhi's second-generation chip products were taped out, Aixin Zhimo AI-ISP was further optimized to significantly improve its performance. AI-ISP can take clear photos of surrounding objects in darker environments, including intersections, mountains, ponds in the dark, and the shadows of people walking and running. Aixin Yuanzhi took nearly three years to bring this technology to market.

After solving the basic problems in the first generation of chips,Aixin Yuanzhi focused on cost-effectiveness in its second-generation chips."For example, how much additional cost can manufacturers accept for new features. Many manufacturers have expressed their willingness to cooperate, but they require that the cost be almost not increased while achieving the desired effect," said Qiu Xiaoxin.

Therefore, Aixin Yuanzhi continues to hone its technology and products, and continuously improve the technical performance of the NPU in its chips. Aixin Yuanzhi's NPU technology reduces the obstacles of memory and power consumption walls to a certain extent by reducing data transfer in response to the needs of manufacturers, thereby supporting more intelligent algorithms with higher effective computing power when the edge side area and power consumption are limited. Driven by the manufacturers' low-price requirements, Aixin Yuanzhi's chip products equipped with AI-ISP and NPU have been widely accepted by the market.

Although Aixin Yuanzhi is still in the stage of R&D and market expansion, after 5 years of polishing, Aixin Yuanzhi's products now rank among the top in the market share of smart cities and AIoT.Among domestic automotive intelligent driving SoC suppliers, its shipment volume has jumped to second place.

In the next five years, Aixin Yuanzhi will shift its focus to commercialization, seeking higher market share and corporate profits. In Qiu Xiaoxin's view, China's current chip design companies have not yet formed a monopoly, and every company has a chance.

AiXin Yuanzhi is already on the table, and Qiu Xiaoxin is very confident in his company.