Intellifusion launches new generation AI chip DeepEdge10

On November 15, at the opening ceremony of the China Hi-Tech Fair, Intellifusion launched the new generation DeepEdge10.

Dr. Chen Ning, Chairman and CEO of Intellifusion, introduced the DeepEdge10 chip at the opening ceremony of the High-Tech Fair. DeepEdge10 is the first domestically produced 14nm chiplet large model inference chip in China. It adopts independent and controllable domestic technology, contains domestic cores, and supports large model inference deployment. The X5000 inference card built on the innovative D2D chiplet architecture of the self-developed chip DeepEdge10 has been adapted and can carry SAM CV large models, Llama2 and other tens of billions of large model operations, and can be widely used in AIoT edge, mobile and other scenarios.

DeepEdge10 Max Yuntian Lifei is one of the few AI companies in China that has insisted on chip research and development since its inception. It has completed the research and development of the third-generation instruction set architecture and the fourth-generation architecture, and has been commercialized one after another. More importantly, through years of investment, it has established a core chip team with an average of more than 14 years of design experience.

Yuntianlifei's chips have also won special projects from the Ministry of Industry and Information Technology, the National Development and Reform Commission, and the Ministry of Information Technology, and have won the "Wu Wenjun Artificial Intelligence Science and Technology Award" three times.

In the era of big models, AI inference chips are the key carrier for application implementation

The market value of edge computing is showing a rapid development trend. It is predicted that by the end of 2023, the global edge computing market will reach a scale of 200 billion US dollars; by 2026, the edge computing market is expected to exceed 300 billion US dollars.

Edge computing scenarios present the characteristics of fragmented computing power, long tail, non-standardization, and scale fragmentation. Traditional algorithm development and chips are difficult to adapt to the product requirements of the new generation of artificial intelligence edge computing scenarios. The emergence of large models provides the industry with a solution at the algorithm level. However, for large models to play a role in actual combat in edge computing scenarios, they need the support of AI large model inference chips.

For AI chips, large models bring new computing paradigms and computing requirements. Chips need to have greater computing power, greater memory bandwidth, and greater memory capacity to support the operation of massive large models on the edge. At the same time, AI edge inference chips also assume the responsibility of "the last mile of landing applications", which means that AI edge inference chips must not only support AI computing tasks such as large models, but also have strong general computing power.

In response to the above scenario requirements, Intellifusion has created a new generation of edge computing chip DeepEdge10. The chip has SoC master control integration; it uses D2D Chiplet technology and C2C Mesh expansion architecture to achieve flexible expansion of computing power; and it is built-in with Intellifusion's fourth-generation neural network processor.

Based on a series of innovative technologies, Intellifusion has created a series of chips. Currently, three chips have been developed: Edge10C, Edge10 Standard Edition, and Edge10Max; the shipping forms include chips, boards, boxes, acceleration cards, inference servers, etc., which can be widely used in AIoT edge video, mobile robots and other scenarios. The X5000 inference card built on the innovative D2D chiplet architecture of Dee Edge10 has been adapted to and can carry large-scale model operations such as SAM CV large models and Llama2 at the billion level.

At present, Yuntianlifei has provided IP authorization for neural processors to leading domestic AIoT chip design, smart car chip design manufacturers, service robot manufacturers, national key laboratories, etc.

Li Aijun, Vice President of InnoLife and General Manager of Chip Product Line, introduced DeepEdge10 in detail at the InnoLife AI Chip Conference. Algorithm chipization is the core "weapon" of InnoLife to build AI chips.

Since its establishment in 2014, Yuntian Lifei has been committed to independent research and development of chips, and has accumulated the core capability of "algorithm chipization". "Algorithm chipization" is not simply "algorithm + chip", but Yuntian Lifei's AI chip design process that integrates the concepts and ideas of chip designers with algorithms based on its understanding of scenarios and quantitative analysis of key algorithm computing tasks in application scenarios, which can enable AI chips to play a better role in practical applications.

With the support of the core capability of algorithm chipization, Intellifusion has completed the research and development of the third-generation instruction set architecture and the fourth-generation neural network processor architecture, and has gradually put them into commercial use. More importantly, through years of investment, the company has established a core chip team with an average of more than 14 years of design experience.

InnoLife's self-developed chips are also an important engine for the company's self-evolving city strategy. In 2020, InnoLife officially released its self-evolving city intelligence strategy at the China Hi-Tech Fair. The core logic driving the development of self-evolving city intelligence is to create a data flywheel of "application production data, data training algorithms, algorithm definition chips, and chip-scale empowerment applications." Chips are the key carrier that determines the breadth and depth of AI applications, and are also an important computing power support for the construction of self-evolving city intelligence.

In the future, Yuntian Lifei will continue to increase its independent research and development efforts, based on independent control, and use its own "cores" to provide a powerful engine for the development of self-evolving urban intelligent bodies.

Reviewing Editor: Peng Jing