This article is adapted from Lao Shang Kan Technology (WeChat ID: shangkeji)
With the explosive launch of Apple’s iPhone X and Huawei Mate 10, media and capital have increasingly focused on embedded AI chips. On November 3, 2017, a grand industry event—the ThunderWorld 2017 Embedded Artificial Intelligence Technology Forum, gathering top AI technology experts from around the world, was held at the Beijing International Conference Center.
One of the most impressive aspects of this forum was that many AI technology leaders and executives, despite being at the forefront of the industry, were able to calmly recognize the problems and challenges facing the embedded AI industry. The keynote speakers rarely discussed achievements and successes but instead confronted the urgent problems and difficulties that need to be solved in the development of embedded AI, which can be considered a refreshing perspective in a technology forum.
We are about to enter an era of interconnected devices, and AI can be divided into cloud AI and embedded AI. Traditionally, cloud AI, which is more suitable for algorithm training, has attracted more attention. So, why is there still a vast market space and commercial value for embedded AI in the face of more powerful cloud AI solutions?
Sun Li, Vice President of Zhongke Chuangda
During his keynote speech, Sun Li provided several examples: a Boeing 787 passenger aircraft generates data at a rate of 5GB per second, and an autonomous vehicle generates data at a rate of 1GB per minute. The volume of data far exceeds existing network bandwidth. In applications requiring large amounts of data with real-time processing and response, cloud AI struggles to meet demand. For instance, when using a smartphone to capture typical scenes like a sunset or a blue sky, or when trying to capture a moment in sports where the optimal shooting time may only last a few seconds, it is impractical to upload data to the cloud for processing and then download it back locally. These real-world scenarios that require rapid decision-making and execution are actually more suitable for embedded AI. As smart home technologies become more popular, such as intelligent sensing lighting systems that need to use image recognition to determine if users need light, users are also increasingly concerned about data security. They may not want cameras to upload many private images to the cloud, so in such data-sensitive applications, embedded AI solutions are more appropriate.
Sun Li also candidly stated that embedded AI faces several challenges and shortcomings: first, limited computing power; second, high requirements for power consumption control; and third, algorithms need optimization. Additionally, issues such as cost and business models exist. Sun Li pointed out that AI itself cannot be sold; it must be integrated into every scenario, combined with hardware and sensors, along with cloud and terminal solutions, to meet a specific customer need or enhance user experience, ultimately achieving its business model. Overall, cloud AI and embedded AI will coexist and complement each other.
Sun Gang, Global Vice President of Qualcomm
Sun Gang, during his keynote speech, also provided an insightful discussion on the differences and synergies between cloud AI and embedded AI: Artificial intelligence requires vast amounts of data, so AI initially often started from the cloud. However, over time, due to security and user-friendly demands, many applications will gradually migrate from the cloud to the terminal. The most common AI application model in the future may be training and education in the cloud, followed by execution at the terminal.
Tang Wenbin, CTO of Megvii Technology
Tang Wenbin stated that for AI applications, the trend of “edge + cloud” is already very clear. We hope to perform some calculations on the device side to relieve the pressure on the cloud; the device side can provide much faster immediate response capabilities. As more data converges to the cloud, enabling cloud AI to have large-scale data mining capabilities, “cloud + edge” is the superior AI combination solution. Currently, embedded AI solutions on the device side still have many areas to explore, whether in terms of algorithms, chips, or applications, so industry collaboration is needed to advance AI applications significantly.
In fact, the development of AI has many similarities with the early development of computers. Some scientists predicted that humanity could meet all computing needs with just five supercomputers. Some companies have attempted to launch network computers, but due to issues like network transmission capacity and task response time, personal computers and local servers dominated for a long time. However, with the improvement of network bandwidth and technological upgrades, cloud computing has gradually returned to people’s vision due to its excellent cost advantages.
Geng Zengqiang, CEO of Zhongke Chuangda (center)
Geng Zengqiang stated in an interview that in the history of technological development, the popularization of many technologies follows a spiral upward trend. The process of AI popularization will be similar. Although when a new technology appears, it may not be perfect, as long as it is provided with a good application platform, regardless of how primitive or poorly designed the tools are, driven by user demand, the industry will inevitably mature rapidly. The development of embedded AI has just begun, and many challenges need to be collectively addressed by the industry, but this also provides a broad development space for innovators, which is an essential reason why new participants with energy and vitality continue to emerge in the AI field.
From another perspective, many application developers now hope to incorporate AI elements into their applications to make their services smarter and more competitive. However, with the current variety of AI hardware, software, and development tools, how can ordinary developers use AI as a common tool and infrastructure for innovation, just like using water, electricity, and gas?
In response to this issue, Sun Li, Vice President of Zhongke Chuangda, stated in an exclusive interview that regarding the standardization of application tools, we are not worried. Frankly speaking, historically, this will happen naturally, just like the popularization of GPU heterogeneous computing applications. During this process, one aspect is driven by technology, which is one of the reasons we organized this Embedded AI Technology Summit, hoping to recommend the establishment and implementation of industry standards through closer communication. From another perspective, we need to consider issues from the programmers’ perspective, including platform and ecosystem construction, and strive to provide programmers with more convenient AI development tools. We realize that AI development toolkits need to be more intuitive and user-friendly during algorithm training. After completing algorithm training, the model should be easily integrated into mobile devices with a clear architecture. Open tools should have clear labeling services and a developer version upgrade model that can quickly absorb programmer feedback and iterate rapidly. In these areas, Zhongke Chuangda will work with industry partners to make AI development tools more user-friendly and efficient.
Although smartphones are an essential scene for embedded AI applications, the application scenarios for embedded AI are exceptionally vast. For instance, the currently popular unmanned supermarkets are very suitable for using embedded AI solutions for identity recognition and payment verification. In the intelligent manufacturing industry, embedded AI can diligently perform industrial product inspection and quality control. In modern agriculture, drones combined with embedded AI can perform precise monitoring of crop growth…
Despite the many issues that still need to be resolved in embedded AI, such as improving chip performance, optimizing algorithms, controlling power consumption, and exploring business models, I firmly believe that solutions will always outnumber problems. As long as embedded AI can provide users with tangible convenience and efficiency, under the impetus of commercial laws, these issues can gradually be resolved through iterative upgrades.
Building a healthy and sustainable AI ecosystem has become an inevitable choice for AI development. Geng Zengqiang, CEO of Zhongke Chuangda, stated in an interview: The future competition in the AI industry will undoubtedly be a competition of ecosystems and industrial chains. Only by pursuing success in the ecosystem can companies within the ecosystem succeed. For the AI industry to achieve faster and better development, the entire AI industrial ecosystem, from chips to operating systems to algorithms and tools, needs to be integrated. Zhongke Chuangda is attempting to build AI platforms and ecosystems, but we also realize that such platforms and ecosystems cannot be built by Zhongke Chuangda alone. Therefore, we invited many excellent companies, research institutions, universities, and technical experts from home and abroad to this technology forum, hoping to build consensus, promote industry collaboration, and help the entire ecosystem grow better. With the joint efforts of numerous partners, progress in AI ecosystem construction is moving quickly, and more news on AI ecosystem construction will be announced soon. Zhongke Chuangda also welcomes more partners to join in building the AI ecosystem. There will never be a perfect AI, only an AI that continues to progress.
We are striving to create a rich and colorful intelligent world,
Your attention and support are the greatest encouragement to us.
