The Yunzhisheng AI technology team led by Chairman/CTO Liang Jiaen of Yunzhisheng brought full of technical sharing. At the same time, the open day also set up two sub-forums "Intelligence + Promote the Integrated Development of the Real Economy" and "Artificial Intelligence Chip

June 5, the 2019 Yunzhisheng AI Technology Open Day (Open Day) opened grandly in Xiamen. Xiamen Station is the fourth stop after Beijing, Shenzhen and Shanghai, and it is also the final stop of this open day.

The Yunzhisheng AI technology team led by Liang Jiaen, chairman of Yunzhisheng/CTO, brought full of technical sharing. At the same time, the open day also set up two sub-forums: "Intelligence + Promote the Integrated Development of the Real Economy" and "Artificial Intelligence Chip Technology and Industrial Application". Combined with industrial practice cases, the current industry is concerned about in-depth interpretation of the implementation of AI chips and AI applications.

production and research combination, Yunzhisheng Xiamen Artificial Intelligence Supercomputing Platform unveiled

, founder/CEO of Yunzhisheng, said that artificial intelligence is a cutting-edge technology that integrates algorithms, computing power and big data. Yunzhisheng has leading algorithms, the ability to condense algorithms on chips, the ability to supercomputing platform, and the ability to commercialize technology. In the future, Yunzhisheng will actively participate in the artificial intelligence development plan of Xiamen and Fujian Province, and conduct extensive cooperation with local enterprises in the fields of smart home, smart medical care, smart government, etc., provide the industry with leading artificial intelligence technology and promote the implementation of technology and industry.

Under the witness of the government, institutions and industry leaders of Xiamen and Fujian Province, Yunzhisheng also officially held a strategic cooperation signing ceremony with pillar state-owned enterprises such as China Telecom Xiamen Branch, Xiamen Radio and Television Group , Xiamen Municipal Group Municipal Development Company, Quanzhou Transportation Development Group, and officially unveiled the plaque of Yunzhisheng Xiamen Artificial Intelligence Supercomputing Platform and Yunmaohuo United Capital Company.

This year, Yunzhisheng Xiamen artificial intelligence supercomputing platform will achieve floating-point computing capabilities of 100 million times per second, which can support the development of artificial intelligence services in many fields such as government affairs, medical care, education, finance, transportation, security, and intelligent customer service. It is a benchmark work of domestic artificial intelligence supercomputing.

Yunmao Internet will rely on Yunzhisheng's industry-leading artificial intelligence technology and Shimao Group's diversified industrial resources to build a "platform-level" comprehensive solution, create full-scene AI applications covering the real estate industry, promote the intelligent upgrade of Shimao Group's entire industry, and serve the construction needs of smart cities.

Full stack AI hard-core technology promotes the second half of the Internet

Yunzhisheng Chairman/CTO Dr. Liang Jiaen first reviewed the three waves of development of AI and pointed out that with the arrival of the third wave of AI, the technical foundation, industrial foundation and demand are becoming increasingly mature, and industrial forces have become the main force leading industry changes. When AI returns to rationality, behind technology is a great era of smart Internet of Things industry transformation, which is the second half of the Internet. This industrial transformation requires technology and industrial applications to iterate forward.

Based on this, Yunzhisheng has refined a set of "full stack" and "hard core" AI technology implementation philosophy - to stimulate the value of AI technology to the greatest extent and promote industrialization, it is necessary to have full stack AI hard core technology. On the scene, Dr. Liang Jiaen fully demonstrated the full-stack AI technology map covering perception, cognition and expression of Yunzhisheng, the full-stack AI industry capabilities that run through technology, products and industry closed loops, and the hard-core strength from underlying algorithms to AI chips.

focuses on the two dimensions of "full stack" and "hard core". On the spot, Dr. Li Xiaohan, Vice President and Head of Chip R&D of Yunzhisheng, Sun Jianqing and Liu Qingsong, senior experts of Yunzhisheng AI Labs, gave in-depth explanations of Yunzhisheng AI technology strength from the perspective of Yunzhisheng AI chips, voice synthesis, and supercomputing platforms.

Li Xiaohan pointed out that AI chips occupy a core position in the implementation of AIoT - on the upper hand, AI chips need to undertake cloud-based capabilities device carriers, which are the hardware support for cloud-based implementation; on the lower hand, AI chips provide AI local computing power based on deep learning, and are a support platform for intelligent networking entrances across devices. It is also for this reason that edge AI chips need to break the design idea of ​​traditional chips, define chips with scenarios, and provide multimodal processing capabilities based on the idea of ​​end-to-cloud interaction, achieve an excellent balance in performance, power consumption, and area, and take into account the needs of connection and security.

Yunzhisheng proposed an architecture-level near-computing storage design (short wall) and an accelerator near-computing storage structure, which improves the data handling efficiency by three orders of magnitude. In terms of power consumption, Yunzhisheng has created low-power voice wake-up technology and low-power microISP dedicated to machine vision to meet the low-power needs of edge-side applications. In terms of security, Yunzhisheng improves security performance through confidentiality and integrity security IP planning.

Yunzhisheng AI Labs Senior expert Sun Jianqing shared Yunzhisheng’s solutions in this field based on voice synthesis technology. Speech synthesis, also known as Text to Speech technology, is the "mouth" of the device.

At this stage, the challenges of speech synthesis are mainly in three aspects: high quality, personalization and diversity. Based on deep learning, Yunzhisheng's speech synthesis technology solution has extremely high-quality end-to-end synthesis and high-quality personalized synthesis, which can convert text into natural, smooth and clear voice in real time and accurately, and has rich ability to expand applications (personalized singing synthesis/story synthesis). In the future, speech synthesis will also show the following trends: at the algorithm level, the ultimate optimization of end-to-end methods to solve computing complexity; at the data level, large-scale data learning enriches information + small data personalization; at the application level, speech synthesis will develop towards multimodal and emotional.

With the continuous advancement of AI technology, intelligent computing is also facing new challenges. Supercomputing platforms are the basis for the value of big data and machine learning , and are also the key to efficiently building AI capabilities. Liu Qingsong, a senior expert of Yunzhisheng AI Labs and director of Xiamen Deep Learning Intelligent Engineering Research Institute, explained to us the Yunzhisheng Atlas supercomputing platform.

To achieve efficient and intelligent computing needs for AI application scenarios, Yunzhisheng takes Atlas large-scale computing platform and UniFlow computing framework as the core, and takes machine learning as the technical entry point, providing a friendly visual interface, complete data processing process, rich machine learning algorithms, efficient GPU/CPU heterogeneous computing platform and distributed storage system, as well as automatic adjustment of participating model architecture search, helping to achieve rapid expansion of the AI ​​technology map.

industry empowers, AI technology promotes the intelligence + transformation of traditional industries such as real estate, agriculture, and communications. In the forum, Chen Jisheng, President of Yunzhisheng Innovation Department, and Yao Peng, Senior Expert of Telecom Group 5G, Liu Jun, Deputy General Manager of Shanghai Shimao IoT Technology Co., Ltd., Xue Sujin, co-founder of Nongxin Internet/General Manager of Nongxin Cloud, and Yue Peng, Chief Architect of Xinhua Third Group Smart City Research Institute under Unigroup , combined with the actual cases of their respective companies, shared the key points and difficulties of AI technology in various scenarios.

, President of Yunzhisheng Innovation Business Department, Chen Jisheng, pointed out that under the new round of AI industry transformation, the number of devices has grown on a large scale, the connection costs are cheaper, the data dimensions are complex and diverse, and the application scenarios faced by technology are also more vertical. Therefore, multimodal AI capabilities have become the key to opening up the closed loop of the AIoT industry.

To this end, Yunzhisheng has created a multimodal AI technology research and development system with a "pyramid" structure. At the bottom, there are DeepFlow clusters and AI chips to provide basic computing resources and underlying support platforms. The Atlas platform in the middle provides a distributed machine learning parallel computing platform, and supports AI technologies such as speech recognition (ASR), speech synthesis (TTS), natural language understanding (NLU), machine translation (NMT), voiceprint recognition (VPR), computer vision (CV) and other AI technologies.

Telecom Group's senior 5G expert Yao Peng gave a comprehensive introduction to the development of 5G technology and the exploration of application. Yao Peng believes that 5G is not only an iterative upgrade of communication technology, but also a far-reaching all-round change and the enablement of digital transformation in the entire industry. In the future, the three major scenarios of 5G technology are: enhanced mobile frame (eMBB), including applications such as high-definition video, virtual reality, and augmented reality; massive physical networking (mMTC), including applications such as smart cities and smart homes; low latency and high reliability communication (URLLC), and typical applications include industrial control, drone control, intelligent driving control, etc.

Liu Jun, deputy general manager of Shanghai Shimao IoT Technology Co., Ltd., shared his experience in exploring intelligence + real estate from the perspective of real estate. Based on the technical system of ABCIET (artificial intelligence, big data, cloud computing, IoT technology, edge computing and communication technology), Shimao IoT has built a smart space management and equipment management platform, and exported low-cost and high-efficiency solutions and informatized, intelligent and flexible standard systems to the outside world. At present, ShimaoWool's intelligent + real estate solutions cover three major sectors: smart community, smart office, and smart market, covering multiple scenarios such as home, community, office, building, park, and municipal.

Xue Sujin, co-founder of Rural Credit Internet/General Manager of Rural Credit Cloud, shared from the perspective of smart pig farming, how Rural Credit Internet uses AI and Internet of Things technology to re-empower the traditional pig farming industry. Xue Sujin pointed out that the "Zhu Xiaozhi" automated pig farming solution of Rural Credit Internet is based on artificial intelligence + big data + Internet of Things. Through cloud boxes, bridges and intelligent collection terminals, people, pigs, equipment, and scenes in pig farms are connected together. Through the two main entrances of Zhu Xiaozhi APP + Zhu Xiaozhi backend, it provides pig farms with intelligent environmental control, precise feeding, smart pig pointing, intelligent weighing, intelligent situation checking, intelligent fat checking, intelligent B-super, and real-time supervision system of smart pig farms.

Yue Peng, chief architect of Xinhua Third Group Smart City Research Institute under Unigroup, interprets the future city. Yue Peng believes that future cities will form a future urban pattern of "intelligence + one network, two connections and three times". Therefore, the technical framework of the urban data brain needs to achieve cloudization of the basic equipment layer, intelligentization of the business support layer, and intelligentization of the industry application layer. Based on this, H3H3 has formed a complete new smart city construction capability system based on the "Four Big One Cloud" strategic business segment and a professional and efficient digital business platform, including top-level design and business consulting, project integration and resource integration, product research and development and solution innovation, urban operation and continuous expansion of new smart city construction.

is accelerating its implementation, and the basic implementation of typical AI chip technologies

AI chip sub-forum, Li Xiaohan, Vice President of Yunzhisheng, and Liu Zhizhe, Director of the Micro System Center of Aerospace Tuowei , Deng Boren, Chip Design Director of the Internet of Things Business Department of Aojie Technology (Shanghai) Co., Ltd., Zhou Jie, Business Director of SiFive China, and Wang Zhiheng, Deputy General Manager of Operations of Zhiyuan Technology Co., Ltd., and other guests shared the key technologies, design difficulties and scenario applications of AI chips.

Li Xiaohan pointed out that a main appeal in the AIoT era is to sink cloud capabilities to devices. Therefore, IoT chips will present the following 5 major trends: 1. Turn from general architecture to AI architecture, provide edge computing capabilities under the premise of controllable cost, and achieve better energy efficiency ratio; 2. The design of AI chips should focus on the organic combination of software, hardware, and scenarios; 3. IoT chips should pay more attention to application optimization, and focus should shift from PPA to APP; 4. The interaction is experiencing the trend of evolution from single mode to multimodal; 5. Edge chips need to be effectively combined with the cloud to achieve end-cloud interaction.

Based on this, Yunzhisheng proposed the Skills On Chip technology path, and through the DeepNet2.0 multimodal neural network processor IP, it outputs a multimodal AI integrated software and hardware solution that combines edge capabilities and cloud capabilities.

Liu Zhizhe, director of the Micro System Center of Aerospace Tuowei Technology, shared the system-level product solutions for intelligent hardware. Liu Zhizhe said that the current mainstream artificial intelligence chips can be divided into three categories: brain-like computing, memristor implementation and ASIC implementation deep learning networks. Among them, processor + artificial intelligence accelerator is the mainstream idea for implementing deep learning algorithms. The main technical implementation method is to independently configure each layer of the neural network accelerator, combine software programming to realize any deep learning network model, and then intensive operations such as convolutional layers, pooling layers and fully connected layers are implemented using hardware accelerators.

Deng Boren, chip design director of the Internet of Things business unit of Aojie Technology (Shanghai) shared Aojie Technology's business layout in the 5G+AI+IoT field.With the development of 5G and AI technologies, IoT devices will grow by 17% in the next few years. While AIoT brings huge market opportunities, the industry also faces challenges such as data security, power consumption requirements, and flexible configuration. In response to this, Aojie Technology's IoT chips include key areas such as eMTC, NB-loT, LoRa, GPS, WiFi, and Bluetooth. They also have independent intellectual property rights for multiple protocol stack software, providing customers with complete solutions from core to cloud.

SiFive China Business Director Zhou Jie pointed out that in the post-Moore era, a chip is suitable for all scenarios no longer applicable. Customization is the only way to improve chip performance, and the only criterion for realizing chip customization is application. Therefore, the future chip customization model must be vertical application (DSA). Saifang provides customized solutions for vertical applications. Its RISC-V CPU core has three major features: standard instruction set core, maximum flexibility, and adhere to open source projects. At the same time, the RISC-V CPU IP product classification has clear application areas, which can help enterprises reduce IP and development costs, shorten R&D cycle, reduce risks and improve customization capabilities, reduce the professional knowledge required for multiple categories, and provide a business model of scalable cloud services.

Wang Zhiheng, deputy general manager of operation of Zhiyuan Technology Co., Ltd., said that in the AIoT era, the biggest challenge of AI chips lies in taking into account both. At the scene, Wang Zhiheng shared the successful cases of Zhiyuan Technology’s cooperation with Yunzhisheng.

Zhiyuan has more than 25 years of end-to-end ASIC design service experience, and has accumulated complete CPU enhancement and IP customization technology, which can support ultra-low power operation mechanisms, including dynamic voltage and frequency regulation (DVFS), power consumption mode conversion, and fast system wake-up, and provides SoC software services.