Domestic AI "chip" potential! Kunyun Technology is committed to providing the next generation of artificial intelligence computing platform

The "2021 Hard-core Chinese Chip" selection event hosted by Xin Shiye is in full swing. More than 100 domestic chip companies have signed up to participate in the selection, and more than 200 products will meet with a large number of electronic engineers.

We are now using the "cloud exhibition" to fully display Chinese chip products and enterprises for you. At the same time, please cast your precious vote for the advancement of Chinese chips!

Online voting time

August 15 - September 14

2021 Best domestic AI chip

The Most Innovative IC Design Company of 2021

Kunyun Technology is a technology-leading artificial intelligence chip company, co-founded by Wayne Luk, an international authority in the field of custom computing chips, a fellow of the Royal Academy of Engineering, UK, Dr. Niu Xinyu and Dr. Cai Quanxiong, etc., and is committed to providing the next generation of artificial intelligence computing platforms with high performance, low latency, high computing power and cost-effectiveness. Focusing on the AI ​​computing engine using the CAISA data flow architecture, relying on the technical accumulation of the main creative team for nearly three decades since 1991, Kunyun has successfully launched the world's first data flow AI chip. Its self-developed CAISA architecture supports deep learning general algorithms while achieving a chip utilization rate of up to 95.4%, which is up to 11.6 times higher than similar products in terms of chip utilization, and has achieved a technological breakthrough in the measured computing power of chips.

At present, Kunyun's Starry Sky acceleration card has been implemented in smart cities, smart manufacturing, smart remote sensing, safety and supervision production, autonomous driving and other fields. It provides integrated algorithm and computing power software and hardware solutions for the above vertical fields, and is applied to typical scenarios such as urban security, industrial quality inspection, safe production, and smart remote sensing. It has reached cooperation with Intel, Inspur Information, DELL, China Unicom, Feiteng, Kylin Software and many well-known universities at home and abroad to accelerate the industrialization of AI technology.

CAISA chip

①Product performance:

As the world's first AI chip using data flow technology, the CAISA chip is equipped with four CAISA 3.0 engines, with more than 16,000 MAC (multiply-accumulate) units, a peak performance of 10.9TOPs, and a chip utilization rate of up to 95.4%. In terms of the measured computing power of the chip, with only 1/3 of the peak computing power, the CAISA chip has increased chip utilization by up to 11.6 times compared with similar products, achieving a measured performance of up to 4.12 times that of similar products.

②Price competitiveness:

The CAISA chip uses a self-developed data flow technology architecture to achieve uninterrupted and continuous computing of the CNN algorithm's computing data within the CAISA chip architecture, achieving a technological breakthrough in measured computing power, and can achieve a chip utilization rate of up to 95.4%. Under the same peak computing power conditions, the CAISA chip can achieve more than three times the measured computing power of similar GPU products, and the chip computing power cost-effectiveness is significantly ahead of international mainstream AI chips.

③Technological innovation:

The CAISA chip is based on two independently developed core technologies: the customized data flow architecture CAISA3.0 and the end-to-end compilation tool chain RainBuilder. It has applied for more than 160 patents, covering data storage architecture, data communication network and overall architecture, underlying operator architecture, RB-Compiler, RB-Runtime, RB-Driver, etc., and has a full set of independent intellectual property rights.

Relying on the technical accumulation of the main creative team for nearly 30 years since 1991, Kunyun Technology has developed a customized data flow CAISA architecture, which breaks through the technical limitations of the traditional instruction set architecture through multi-engine parallel computing. Under the same peak computing power conditions, the CAISA architecture can bring up to 10 times more chip utilization.

While the chip architecture provides high computing power, the end-to-end compilation tool chain RainBuilder that comes with the CAISA chip seamlessly links to multiple mainstream frameworks such as Caffe, TensorFlow, PyTorch, ONNX, and supports multiple algorithm models such as VGG, ResNet, and YOLO. It provides high software usability, allows for fast and low-cost algorithm migration and deployment, lowers the threshold for use, and provides better end-to-end performance at the system level.

With its leading technological advantages and scenario-based implementation capabilities, the CAISA chip won the SAIL Award, the highest award at the 2021 World Artificial Intelligence Conference, and has also won honors such as the World Computer Conference Innovative Technology and Product Application Results and the China Hi-Tech Fair Outstanding Product Award.

④Customer Service:

As an edge AI inference chip, the CAISA chip provides deep learning inference at the edge and data center for server manufacturers, algorithm manufacturers, etc. Currently, the Starry Sky acceleration card and AI video analysis platform based on the CAISA chip have been used in petroleum, chemical, electric power, industry and other fields.

⑤Application case introduction:

The series of data stream AI computing platforms equipped with CAISA chips support multi-channel video structured analysis in different computing power scenarios, and provide high-performance, low-latency, high-computing-power and cost-effective computing acceleration solutions for security, safety supervision, oil fields, intelligent manufacturing and other scenarios.

In the oil field, AI acceleration hardware and deep learning algorithms based on CAISA chips can process 300 video data from an oil plant, identify oil leaks, helmets, work clothes, intrusions, and the operating status of oil production equipment, and achieve a time of as fast as 1 second from video stream acquisition to alarm output, providing more reliable safety protection for the production situation in the oil field.

In the chemical industry, the algorithm and computing power integrated solution based on the CAISA chip can realize early warning and prediction of dangerous sources such as smoke and flames for more than 200 video images connected to the emergency response center, providing auxiliary decision-making for emergency handling of accidents in the industrial park.

In the cement industry, AI acceleration hardware and high-precision AI algorithms based on CAISA chips can identify and alarm the ship's loading volume and dust/raising during unloading through the AI ​​visualization monitoring center, assisting operators in completing on-site operations and providing decision makers with a more comprehensive and visualized on-site situation.