A ban on the United States caused a serious impact. The crisis of ZTE's 'core pain' has become a hotly debated topic for the people. As a result, people have realized a reality: The chip is important, but it is very difficult to do.
Internet of Things (IoT, InternetOfThings) Artificial intelligence services companies, the main business is generally to provide a 'cloud' program, and how cloud knowledge will be 'steer to reality' R & D AI chip? This is the common question of the participants in the field, It is also the decision-making root of Yun Zhisheng's team's own request.
'The decision to develop AI chips is the path to light and the sea, but this process is like entering the unknown desert full of hardships and bleakness. During the march in the desert, investors provided food and water, and Yunzhisheng chip R&D The team must personally experience endless torture. 'Yao Zhisheng's founder/CEO Huang Wei said.
Yun Zhisheng Founder/CEO Huang Wei
At the launch event of the first generation UniOne IoT AI chip and its solution, Swift, Zhang Yijia, the founder of Jiazi Lightyear, experienced a change in Huang Wei’s status. She said bluntly, compared with the status of the past three years. , Huang Wei now seems to walk from the night into the daytime, giving people a feeling of sunshine.
Answer: The original chip architecture is difficult to meet the needs of the new situation of AIOT
China National Grid learned that Yun Zhisheng decided in 2014 to independently research and develop AI chips for the Internet of Things. In 2015, he established a chip team to embark on the journey to 'cross the hills'.
Huang Wei pointed out that in the process of the migration and development of the Internet through the PC, Mobile, and AIOT phases, the number of terminal smart devices has grown by orders of magnitude. Not only its form and scenario are ever-changing, but also the fragmentation of demand is serious. The original common architecture chip is difficult to meet new requirements. Situational experience needs.
In addition, the access of the AI algorithm also imposes higher requirements on the parallel computing capabilities and memory bandwidth of the device-side chips. Although traditional GPU-based chips can implement inference algorithms in the terminal, it is not designed for deep learning, and the energy efficiency ratio is not Far less than AI-specific chips.
Huang Wei believes that whether the CPU or GPU, FPGA, the existing chip architecture is not specifically designed for AI, and can not meet the AI computing needs of the Internet of Things, and consider too much backward compatibility, so the performance is far from Optimal. 'Based on the business aspects of the chip product, the repeated verification of the scenario, and the judgment of the AIoT final office, Yun Zhisheng decided in 2014 that it must independently develop AI chips for the Internet of Things.'
Compared with traditional chips, customized AI chips are relatively specific in terms of application scenarios and AI algorithms, so they are more specialized in hardware design, and they have absolute advantages in computing density and power consumption when compared to general-purpose chips. Therefore, compared to the general-purpose chip, the AI chip can perform the calculation of the same task amount in machine learning at a lower frequency and a smaller chip area, so that the multi-dimensional requirements such as cost, power consumption, and calculation power can be achieved. The perfect balance.
Direction: 'Swift' First hand in hand to partner 'Core' Ecology
On May 16, Yun Zhisheng released the first-generation UniOne Internet of Things AI chip and its solution, Swift, in Beijing. The chip was independently designed and developed by Yun Zhisheng and used as a CPU+uDSP+DeepNet architecture. Self-informed AI instruction set, DeepNet with independent intellectual property rights, uDSP (digital signal processor), support for various deep neural network models such as DNN/LSTM/CNN, etc. The performance is 50 times faster than the universal solution.
According to Li Yuhan, co-founder/Chip Leader of Yun Zhisheng, Yun Zhisheng not only provided Swift chips and terminal engines, but also applied the part to the customer for open source, and provided corresponding customization tools and cloud AI capability services. Through cloud core integration, the cloud Zhisheng is based on Swift's turn-key solution for a specific scenario such as smart home, smart speaker, smart car, etc. Based on the Swift solution, customers can stand at a higher starting point for design, at a lower cost. Create more stable and reliable products in a shorter time. At the same time, the open source solution can also ensure that customers can design other long-tail product forms based on the provided AI capabilities and build a richer AIoT ecosystem.
'UniOne is not a chip, but a series of chips, it represents the overall idea of cloud knowledge for the development strategy of the Internet of Things AI chip. ' Li Yu Han pointed out that for solution providers and developers, UniOne can provide a complete reference application of voice AI The solution, cloud capabilities, and customization tools can help customers explore the cost of the IoT hardware product across the minimum time, resources, etc., to create the highest user experience in the future.
Yun Zhisheng Co-founder/Chip Leader Li Yuhan
It is understood that after 'Swift', Yun Zhisheng will launch 'Snow Leopard' smart car chip in 2019 and will launch 'Sailfish' smart city chip in 2020.
At the conference site, Yun Zhisheng announced its strategic cooperation with JD Smart and Gigabit Technology. The parties will be based on the cloud of knowledge in the Internet of Things AI chip and system solutions, sharing their own advantages of resources, build for the new retail, AI 'core ecology' in smart home, automotive and other fields.
In the smart home area, Yun Zhisheng joined hands with Jingdong Intelligence to promote the application of artificial intelligence chips in the smart home field. Together with JD. Alpha Platform, they have collaborated to create customized smart benchmarking products to enable cross-brand, cross-category smart device interoperability and enable users to pass Natural language can get platform music, news, shopping and other massive content services, bringing more intelligent and more convenient and comfortable home life experience.
For Intelligent Connected Vehicles, Yun Zhisheng will work together with Universiad Technologies to jointly develop pre-shipment-class AI chips to create the ultimate in-car intelligent interactive experience and promote the upgrade of artificial intelligence chips in the automotive industry. Building a voice-centric cockpit interaction platform around the GKUI cockpit ecosystem to help users achieve more accurate voice services in terms of location navigation, car entertainment, vehicle control, vehicle-to-vehicle interconnection, etc., let AI empower cars, and make cars It is more 'smart' to let people enjoy smarter travel in a more convenient and efficient way. For the arrival of the smart networked car era, both parties will also carry out AI chip R&D cooperation in the field of automatic driving, and work together for future intelligent travel technology capabilities. The promotion.