Yang Zuoxing, a doctoral graduate from Tsinghua University, the founder of the Shenmu brand, and the chairman of Hangzhou Yanjiwei: Ultra-low power consumption, the future of AI green intelligent applications | 2025 AI Partner Conference
The year 2025 marks the beginning of the explosion of AI applications. As the global AI competition enters the "China Moment," a profound technological revolution is quietly reshaping the industrial landscape. At this critical juncture, the industry faces core questions: How can we bridge the gap between AI technology and large - scale applications? Where will the next disruptive AI super - application be born?
On April 18th, the 2025 AI Partner Conference, hosted by 36Kr, grandly opened at the Modu Space in Shanghai. The theme of this conference is "The Super App is Coming," focusing on the disruptive transformation of AI applications in all industries. The conference is divided into two major chapters: "The Super App is Coming" and "Who Will Be the Next Super App," covering seven major topics such as "Growing Up in the AI World" and "In 2025, Focus on Super Apps in AI." It includes more than 10 keynote speeches, 3 round - table discussions, and two sessions for releasing the lists of excellent AI case enterprises. It deeply analyzes how AI technology reconstructs business logic and reshapes the industrial landscape, and explores the infinite possibilities brought by AI super - applications.
On that day, Yang Zuoxing, a doctor from Tsinghua University, the founder of the Shenmu brand, and the chairman of Hangzhou Yanjimicro, delivered a keynote speech titled "The Future of Extreme Low - Power AI Green Intelligent Applications."
Yang Zuoxing, a doctor from Tsinghua University, the founder of the Shenmu brand, and the chairman of Hangzhou Yanjimicro
The following is the content of Yang Zuoxing's speech, edited by 36Kr:
Dear guests, colleagues, and friends, good afternoon!
After a whole day of intellectual exchanges, I believe you all feel both fulfilled and exhausted. I am very grateful to 36Kr for their kind invitation, which gives me the opportunity to discuss cutting - edge issues in the field of AI with you at this highly significant time. Today, we are gathered here to jointly search for "the next super APP in AI." Coincidentally, DeepSeek has brought a world - leading large - model foundation, which not only achieves performance breakthroughs but also reduces costs by an order of magnitude. Meanwhile, its open - source approach has greatly lowered the threshold for AI development. From the large - model foundation to AI Infra (infrastructure) and then to the chip layer, the key links in the technology chain have gradually been completed, just like "Good rain knows its season, it comes when spring is here."
I am Yang Zuoxing from Hangzhou Yanjimicro. Today, I would like to share with you an important but often overlooked area in the wave of AI technology - low - power technology. It is like the "invisible engine" of AI applications, profoundly affecting the sustainability and application boundaries of AI development.
Looking back at the three development waves of AI, from 2005 to 2015, deep learning first exceeded human - level performance in the field of face recognition, opening a new era of AI applications. After 2013, large models emerged, especially since 2023, the rapid progress of AI technology has given us a glimmer of hope of approaching human intelligence. However, as the model scale grows exponentially, from the early face - recognition models of a few hundred megabytes to today's large models of hundreds of gigabytes, the demand for computing power has increased synchronously. This has directly led to an exponential increase in energy consumption. In 2022, the power consumption of global data centers accounted for 3% of the total electricity consumption. As Elon Musk said, "The end of AI is electricity." Power supply has become a key bottleneck restricting the development of AI.
In addition to power consumption, the heat - dissipation problem is also very tricky. The heat generated by high - power consumption needs to be dissipated through radiators, but the production capacity of aluminum and steel, the main materials of radiators, is difficult to meet the exponentially growing demand. How can we break the vicious cycle of "growing computing power - rising energy consumption - limited resources"? DeepSeek has achieved a balance between "top - level performance and low resource consumption" by optimizing the model and deployment plan, providing a new idea of "increasing revenue and reducing expenditure" for the industry. On the hardware level, reducing chip power consumption has become another key path.
There are mainly two conventional paths to reduce chip power consumption. One is to rely on Moore's Law, reducing power consumption through the progress of process technology from the early 250 - nanometer to today's 2 - nanometer. However, in recent years, the power - consumption reduction rate brought by each generation of process technology has decreased from 60% to about 30%, and the dividends of Moore's Law are gradually fading. The other is to innovate the design method, but traditional low - power design methods, such as gating and voltage regulation, have been used for decades, and technological innovation has reached a bottleneck.
Since 2008, Yanjimicro has been deeply involved in the chip field. From RFID technology to the research and development of computing power chips, it has gradually explored a set of disruptive full - custom chip design methodologies and first shared this concept with the industry at the Xiamen Semiconductor Summit in 2018. This methodology abandons the standard cell library and designs dedicated cells according to specific needs. It abandons the automated generation of netlists and returns to manually writing netlists. It gives up EDA automatic placement and uses manual placement. This full - custom chip design methodology has brought significant performance improvements. By using dynamic logic design, although dynamic logic is error - prone, Yanjimicro has reduced the power consumption to a fraction of that of static logic by precisely controlling technical details. The layout utilization rate of traditional chips is about 60%, while the full - custom methodology can achieve an ultra - high utilization rate of 97%, saving 30% - 40% of power consumption. The manually written netlist achieves 2 - 3 times efficiency optimization at the logical code level. Overall, the full - custom chip design methodology has achieved more than 10 times improvement in area and power consumption. Taking the computing power chip as an example, the power consumption of the traditional solution per teraflop of computing power is more than 500 watts, and the cost exceeds 1,000 yuan. In contrast, Yanjimicro's solution reduces the power consumption to 180 watts per teraflop and compresses the cost to 240 yuan per teraflop, with an overall optimization amplitude of up to 11 times.
Based on the low - power chips of the full - custom methodology, Yanjimicro has achieved innovative implementation of AI applications in multiple fields, with the most representative being the Shenmu brand series of products. This series solves the two pain points of short battery life and high heat generation of traditional AI chips through low - power technology, making AI computing power truly "affordable and easy to use."
The chips used in the Shenmu AI intelligent camera series are equipped with an AI ISP module with 6T computing power. When calling 1.5T computing power, it only consumes a few dozen milliwatts of power. This feature enables it to achieve a leap - forward breakthrough in night - time imaging performance. Ordinary mobile phones or action cameras often produce blurry and noisy images at night, while the Shenmu AI intelligent camera can achieve full - color imaging in low - light conditions. Even in the dim environment at 11 p.m., it can present clear images as if it were daytime. From now on, "the darkness of the night" is no longer an obstacle to visual perception, and AI technology has endowed the camera with all - weather "visual ability."
The Shenmu AI intelligent camera product integrates a general - purpose NPU module with 4T computing power, and the power consumption at 1T computing power is also as low as a few dozen milliwatts. Based on the low - power advantage, the product has realized a rich set of AI detection and recognition functions, supporting the detection of more than 10 types of targets such as people, vehicles, animals, fireworks, and packages, with a maximum detection distance of up to 50 meters and a very low false - recognition rate. It covers functions such as face recognition, vehicle recognition, gesture recognition, and voice command recognition. Users can perform operations such as "call for help," "make a phone call," and "privacy shielding" through voice commands. It can also judge a child's sitting posture and mobile - phone - playing behavior in real - time, recognize facial expressions, and even monitor an infant's sleeping state to prevent the risk of falling.
The most breakthrough function of the Shenmu series is the "call for help" system. Traditional cameras are only used as post - event monitoring tools, while Shenmu, through AI voice recognition technology, will immediately activate a triple - alarm mechanism once it detects keywords such as "call for help." It will first call the preset family number. If the call is not answered, it will automatically initiate a video call. Finally, it will contact the 400 customer service center to report the emergency. This function upgrades the security attribute of the camera from "passive recording" to "active rescue," truly realizing the in - depth integration of technology and humanistic care. In response to the needs of rural left - behind children and empty - nesters, the Shenmu product offers a free family - call function. Children or the elderly only need to say "call mom/son," and the camera will automatically dial the number, breaking the spatial barrier and making family communication cost - free and barrier - free. Just as Yanjimicro's vision states, "Keep family affection from being separated by distance and make technology full of warmth."
Our Shenmu AI has achieved 4A - level perception, which means Anyone, Anytime, Anywhere, Anything. We aim to enable anyone to perceive anything at any time and any place, of course, under the premise of authorization. The full - color imaging in low - light conditions allows the Shenmu AI intelligent camera to have the same excellent visual ability at night as during the day, and the wide coverage of the 4G network makes the application scenarios of the Shenmu AI intelligent camera no longer limited. More importantly, we have reduced the installation and usage costs of the Shenmu AI intelligent camera to zero, and the price has always been kept at an affordable level. This design is to enable everyone to easily deploy the camera wherever they want, truly realizing that everyone can perceive anything in the world at any time and any place.
In terms of top - level protection, in addition to the Shenmu call - for - help function mentioned above, the full - color imaging technology in low - light conditions is also of great significance. Many improper behaviors often occur in hidden environments at night. When the Shenmu camera makes the night as bright as day, potential malicious intentions will naturally be greatly curbed, which undoubtedly significantly improves the level of security protection. Combined with functions such as monitoring whether a baby is about to fall off the bed and firework detection, Shenmu has built a comprehensive and all - around protection system.
We always adhere to the concept of "keeping family affection online in real - time." In modern society, spatial distance often makes the connection between relatives become distant. Even if they care about each other, they may not be able to communicate every day. The combination of the Shenmu camera and AI technology has broken this barrier. The free family - call function makes communication easily accessible. Technology is no longer just cold codes and data but a bond carrying emotions, enabling family affection to cross distances and stay online all the time.
From achieving 4A - level perception, providing top - level protection, to delivering family - affection care, our technology and AI have truly given the product the power to warm people's hearts, allowing everyone to truly enjoy the great changes in life brought by AI. Our company focuses on chip research and development, and the low - power chips we develop have world - leading advantages in balancing performance and power consumption. We sincerely hope to cooperate with more partners in the ecosystem, so that our chips can inject stronger power into AI and jointly explore more imaginative application scenarios.
Yanjimicro always takes "Achieving a better life with excellence" as its mission. We firmly believe that low - power chips are not only the product of technological breakthroughs but also the key to promoting the popularization of AI. From the large - model foundation to AI Infra, and then to chips and applications, every link in the entire AI ecosystem requires the pursuit of "excellence" in technology. Here, we sincerely invite more ecological partners to join hands with us and integrate the advantages of low - power chips into more AI application scenarios. Let's use technology as the pen and innovation as the ink to jointly draw a beautiful blueprint for AI to empower life! Thank you all!
About Hangzhou Yanjimicro
Hangzhou Yanjimicro Electronics Co., Ltd. is an innovative technology enterprise focusing on high - performance integrated circuit design and solutions. It was established in 2020. Led by Dr. Yang Zuoxing, it pioneered the full - custom methodology technology path, revolutionized the chip design process, and has been deeply involved in the fields of analog chips, power devices, sensors, and camera chips. It is committed to promoting the upgrading of the intelligent industry through the full - chain service of "chip design + system integration." The "Shenmu AI intelligent camera" series under its banner relies on independent core technologies, integrates AI algorithms and full - custom chip architectures, and has the capabilities of face recognition, behavior analysis, and ultra - low - power edge computing. It is widely used in scenarios such as intelligent vision, security protection, and intelligent companionship. With its high reliability and high energy - efficiency ratio, the product has quickly gained market recognition and become a new focus in the industry.