Yizhi Electronics brought its self-developed NPU to the 2021 Shenzhen Security Expo

On December 26, the 18th China International Social Public Security Expo was grandly opened in Shenzhen.

As a end-side artificial intelligence chip company that Chinese mainland universal computing power, Yizhi Electronic Technology Co., Ltd. (hereinafter referred to as "Yizhi Electronics") focused on the three product lines of smart security, intelligent in-vehicle and intelligent hardware (AIoT) at this exhibition. Among them, in the field of smart security and intelligent vehicles, it is mainly chip + algorithm (scene) packaging and delivery; the main delivery in the AIoT field is chip + tool chain. "Customers use our toolchain for AI algorithm design, and in the process, EZ Electronics supports customers through toolchains." Chen Feng, founder and CEO of Yizhi Electronics, said.

Self-developed NPU is the only way out for AI computing companies

"Trillions of equipment intelligence", this is the Slogan of Yizhi Electronics, and it is also the company's insistence on the intelligent track.

As early as two or three years ago, the development of the AI industry was in full swing, and the increasingly mature artificial intelligence technology continued to promote the growth of the AI chip market, and the Prospective Industry Research Institute predicted that the global artificial intelligence chip market size would reach 72.6 billion US dollars in 2025. At the same time, China has a huge downstream market, and with the blessing of policies and capital, a number of excellent AI chip manufacturers have emerged.

However, as of now, due to the difficulty of achieving large-scale landing, the bursting of the AI bubble has also made many artificial intelligence companies gradually disappear. At present, Yizhi can be said to be the closest to success. Chen Feng pointed out that as the only enterprise that Chinese mainland large-scale landing, Yizhi's shipments this year have reached millions of pieces, and it is expected to achieve a turnover of more than 100 million yuan this year and a pure chip income of hundreds of millions of yuan next year.

It is reported that Yizhi Electronics was registered and established in Zhuhai in 2016, and currently has branches in Beijing, Shenzhen, Shanghai and Hong Kong, and is a national high-tech enterprise and a unicorn potential enterprise in Zhuhai, and has successively obtained strategic investment from famous venture capital northern light venture capital, Intel Capital, China State Construction Investment and other institutions. In 2020 and 2021, E-Wisdom Electronics was twice selected as the Global EE Times Silicon 100 (100 new semiconductor companies to watch in the world).

Up to now, Yizhi Electronics has completed the large-scale mass production of the second generation of AI chips, while continuing to iterate chip products around the application scenarios, in 2021 has launched high-performance security AI SoC chips SV826 and SV823/SV823C, smart car chips SA226/SA223, AIoT SoC chip SH516, respectively, used in front-end intelligent cameras, intelligent driving recorders, intelligent video department standard machines, AI scanning pens, intelligent PTZ and other products.

Yizhi Electronics is not a traditional security manufacturer, and it is not a traditional security product. Compared with traditional chip manufacturers, the advantage of Yizhi Electronics is its intelligent ability. For intelligent security scenarios, SV826 and SV823/SV823C adopt intelligent H.265+ encoding technology to support up to 4K ultra-high-definition video recording; at the same time, integrate professional security level ISP, support 2-3 frame width dynamic fusion and adaptive noise reduction, and perform well in backlit and low-light environments; and equipped with the second generation of EChi's self-developed NPU, providing 1.5T/1.0T intelligent computing power.

SV826/SV823/823C can be combined with a variety of third-party algorithm applications, "machine non-human" detection and identification, high-altitude projectiles, electric vehicle recognition, helmet recognition and other object recognition algorithms to help intelligent security applications land.

Compared with friendly businessmen, in the field of security, Yizhi Electronics mainly focuses on the improvement of intelligent penetration. Compared with most friends who buy third-party NPUs, the core NPU IP of Yizhi is completely independently developed. In response to this, Chen Feng put forward a question that the industry is very concerned about: Why can self-developed NPUs make the market a blue ocean? "Mainly because GPUs have industry standards, which means that all companies in the entire ecological chain can perform their duties, but artificial intelligence has not yet had an industry standard for NPUs," he explained later.

"At this point in time today, only self-developed NPUs are the only way out for AI computing companies."

For example, in the absence of industry standards, the third-party IP purchased by the company means that the support of AI scenarios must be case by case, that is, it cannot be accumulated. Therefore, at present, there are no enterprises that can achieve large-scale mass production by purchasing third-party IP.

At the same time, in addition to self-developed NPU, based on more than 20 years of stable industrial background, Yizhi Electronics for customer needs and understanding, delivery, product strength and service capabilities are also in a leading position in the industry.

"Chip company is ultimately an ecological company, so at this point in time today, we have not sold any board or module, and at present, based on the platform of Yizhi, design-in has more than 200 customers, and design-win customers have approached 100." Chen Feng said.

Continuous iteration of the mainstream hashrate interval

In recent years, the localization of chips has been replaced, and the safety and control have been carried out in full swing, for which Chen Feng believes that in the Chinese mainland, the future demand for Xinchuang (localization substitution) should be more and more intense. And Yizhi's SoC chip, not only the chip has achieved localization substitution, the technology and IP inside have basically achieved localization.

"We are not only a localization chip company, the vast majority of technology and IP have also reached the degree of localization."

In view of the difficulties of product landing, Chen Feng frankly said that more exist in the commercial field. Since AI is still an emerging thing, for customers, the acceptance of emerging fields and the support of technology are the first difficulties that need to be overcome.

Relatively speaking, AI shows a relatively fragmented phenomenon in the field of intelligent hardware, which means that customers' support for original requirements is also relatively high. "The consistency of the algorithms of customers and third-party algorithm companies is relatively low, so the proportion of support for different algorithm manufacturers is larger."

Next, Chen Feng analyzed from the technical level, first of all, from the customer's point of view, it is best to solve all problems with one chip, which means that the mass production is the highest. Because the more chips, the more complex the system, based on the principle that the manufacturing yield is inversely proportional to the chip solder joint, the product manufacturability is poor; at the same time, the increase in chips will also lead to poor product testability and maintainability.

Looking forward to the future, based on the changing characteristics of the mainstream hashrate range driven by different time points and different needs, Yizhi will continue to iterate the hash rate range of mainstream computing power. "This year's mainstream computing power may be 0.5-1.5T, maybe in a year or two it will be 2-3T, so we will continue to iterate NPU and ISP technology, especially in the field of ISP technology, I believe that the AI ISP technology developed by the company will provide better image quality." Of course, we have 3 years of actual scale mass production experience in NPU, and we will continue to iterate on better architectures. Chen Feng revealed.