On April 29, 2022, Cambrian-U (688256) announced that Chuangjin Hexin Fund, Tamsui Spring (Beijing) Investment, Goldman Sachs (Asia) Co., Ltd., ICBC Credit Suisse Fund, Everbright Prudential Fund, Everbright Asset Management, Guohua Life, Guotai Junan, Hengqin Life, Huachuang Asset Management, Huafu Fund, Huatai Securities, Huatai Asset Management, HSBC Jinxin, Huisheng Fund, Chaos Investment, BOCOM Schroder King, Lenovo (Beijing), Lingwei Private Equity Fund, Minsheng Securities, Mingshi Partnership Fund, Mingze Investment , Panjing Investment, Pengyang Fund, Qinmu Assets, Rongtong Global Investment Limited, Ruida Fund, Shanghai Jisheng Investment, BANK OFC Fund, Tianhong Fund, Western Lide Fund, Cinda Aoya Fund, Cinda Australia Bank Fund, Xinhua Fund, Industrial Fund, Industrial Securities Global Fund, Yongying Fund, China Overseas Fund, BOC International Asset Management, BANK OF China Fund, china Post Life Insurance surveyed our company on April 7, 2022.
The main contents of this survey are:
Q: Please summarize the new research and development progress or achievements made by the company in 2021.
A: In 2021, the company will continue to increase its product research and development efforts. In terms of hardware, the company released a training integrated Siyuan 370 smart chip and acceleration card based on the fourth generation of intelligent processor microarchitecture (MLUarch03). Siyuan 370 is the third generation of Cambrian cloud products, using TSMC 7nm advanced process technology, the maximum computing power is as high as 256TOPS (INT8), which is twice the computing power of The second generation of Cambrian cloud reasoning product Siyuan 270. At the same time, the Siyuan 370 smart chip adopts advanced Chiplet core particle technology, which supports the flexible combination between the core particles, and achieves the commercial use of a variety of intelligent acceleration card products with only a single tape-out. At present, the company has launched 3 acceleration cards: MLU370-S4, MLU370-X4, MLU370-X8, and has carried out in-depth application adaptation with domestic mainstream Internet manufacturers. In terms of software, the company has invested a lot of resources in optimizing the basic system software platform, and the unified software platform is improving day by day. In 2021, the company released a new inference acceleration engine, MagicMind. Relying on the downstream components of the Cambrian basic system software platform, MagicMind has improved the debugging and optimization toolchain, and continuously improved the open optimization and debugging capabilities. With magicMind on algorithm models trained on MLUs, GPUs, and CPUs, users can deploy inference services to a full range of Cambrian products with minimal development costs and obtain competitive performance. At the same time, the company actively embraces the open source ecosystem and develops a training software stack that combines high performance and versatility. Native support for the industry's open source frameworks Pytorch and Tensorflow, both of which provide perfect infrastructure support, users can quickly complete the migration of model code based on open source frameworks. In the field of distributed systems, the company has developed a multi-machine and multi-card interactive acceleration library CNCL, and has carried out deep adaptation on open source training frameworks such as Pytorch and Tensorflow, supporting data parallel training of conventional networks and parallel training of NLP large models. The Pytorch and Tensorflow frameworks that are deeply adapted to CNCL can achieve sublinear acceleration of 100-card large-scale data parallel training without modifying the training code, realize parallel training of tens of billions of networks, and give full play to the advantages of computing power to complete large-scale training services. In addition, the company set up a holding subsidiary Xingge Technology to carry out the research and development and productization of on-board intelligent chips. According to the differentiated demand for artificial intelligence computing power in the automotive market, Xingge Technology has planned vehicle intelligent chip products in different gears. The research and development of a new generation of products and in-vehicle intelligent chips is being carried out in an orderly manner, and the company's strategy of coordinated development of "cloud-edge end vehicles" has been comprehensively promoted.
Q: Please introduce the company's new cloud training acceleration card MLU370-X8.
A: The company's newly launched training acceleration card MLU370-X8 is a dual-slot 250W full-size acceleration card, which can provide 24TFLOPS (FP32) training computing power and 256TOPS (INT8) inference computing power, while providing FP16, BF16 and other training accuracy. MLU370-X8 is equipped with dual-chip four-core grain for the first time, providing twice the memory, codec resources of the standard Siyuan 370 accelerator card, and equipped with MLU-Link? multi-core interconnection technology, each acceleration card can obtain 200GB/s communication throughput performance, which is 3.1 times the bandwidth of PCIe4.0, mainly for training tasks with higher computing power and bandwidth requirements.
Q: The company's edge smart chips and acceleration cards will achieve considerable sales revenue in 2021, what are the main customer groups for this product line? What is the difference between it and the cloud smart chip and acceleration card?
A: Edge computing is a new computing paradigm that has emerged in recent years, equipped with moderate computing power on the device between the terminal and the cloud, which can effectively make up for the disadvantage of insufficient computing power of the terminal device on the one hand, and alleviate the potential problems such as data privacy, bandwidth and latency in cloud computing scenarios on the other hand. Siyuan 220 chip and its acceleration card is the company's first edge intelligent chip product, in 2021 landed a number of head enterprises, to achieve nearly one million pieces of large-scale sales, can support the edge computing scenario under the intelligent data analysis and modeling, vision, speech, natural language processing and other diversified artificial intelligence applications, to promote intelligent manufacturing, intelligent retail, intelligent education, smart home, smart grid and many other fields of rapid development.
Q: What is the specific situation of the company's self-developed basic system software platform upgrade and optimization?
A: Cambrian basic system software platform is a software development platform specially built by the company for its cloud, edge and end intelligent processor products, using a cloud edge end integration, training and promotion integration architecture, which can support the cambrian cloud, edge and end of the full range of products at the same time. The company's smart chip products share the same software interface and complete ecosystem, which can easily develop, migrate and optimize intelligent applications. The company's basic system software platform mainly includes training software platform and inference software platform. The training software platform supports rich graphics, voice, recommendations, and hyperscale NLP training tasks. Through the underlying high-performance operator library CNNL and high-performance communication library CNCL, the industry's leading hardware computing efficiency and communication efficiency are achieved in the actual training service, and the rapid migration method of the model is provided to help users quickly complete the migration of the existing business model. For the inference software platform, the company added a new inference acceleration engine MagicMind, which is the industry's first inference engine based on MLIR graph compilation technology to achieve commercial deployment capabilities. With MagicMind, users can deploy their inference business to the full cambrian range of products with minimal development cost and competitive performance.
Q: What is the planning and progress of the company's on-board smart chips?
A: Xingge Technology, a holding subsidiary of the company, is developing and designing different levels of automotive intelligent chip products. It takes a certain period of time for a chip to be designed and finally mass-produced. On the one hand, based on the company's judgment of the technology route and the mainstream trend of the industry, Xingge Technology will design and develop with the help of R&D accumulation in the field of cloud smart chips. On the other hand, according to the automotive industry itself pays more attention to functionality, security and the adaptability of the software platform, Xingge Technology also needs to superimpose and design chips that meet the requirements of the vehicle regulations on the basis of the existing chip technology components to help the company build a unified intelligent ecology of "cloud-edge end vehicles".
Q: According to the company's "2021 Annual Report", the revenue for the whole year of 2021 was 721 million yuan, an increase of 57.12% compared with the same period last year.
A: In 2021, the company's market expansion has made certain breakthroughs, and its operating income has grown rapidly, an increase of 57.12% over the same period in 2020. Among them, the company's edge-end product Siyuan 220 chip and acceleration card landed in a number of head enterprises, in 2021 to achieve nearly one million pieces of large-scale sales, edge product line revenue increased by 741.10% over the same period last year; the company's cloud product line business steadily advanced, training the whole machine also made certain sales progress; the company's intelligent computing cluster system business relying on the benchmarking effect and excellent reputation of the previous project, the revenue in 2021 increased by 39.91% over the same period last year.
Q: How does the company view continuous R&D investment?
A: At present, the market competition in the field of artificial intelligence chips is becoming increasingly fierce, which is not only valued by many leading integrated circuit companies, but also becomes the focus of many start-up integrated circuit design companies. Moreover, the integrated circuit design industry is a typical talent-intensive industry, and the R&D team with high professional level and strong technical strength is the guarantee of the company's continuous innovation. Therefore, as a smart chip design company, in order to occupy a certain market position in the fierce market competition, and withstand multiple tests such as time, industry development, and macroeconomic environment, it is necessary to firmly and continuously invest in research and development, which is used for the upgrading and iteration of the underlying basic technology of the chip, the research and development of new products in each product line, the introduction of outstanding talents and the stability of the core team, so as to enhance the core competitiveness of products in an all-round way and better serve customer needs. In 2021, in the fiercely competitive market, the company has achieved steady progress in product research and development, market expansion, ecological construction and other aspects, which has comprehensively guaranteed the steady development of the company. The company released the 370 smart chip and acceleration card of the integrated siyuan, which showed certain advantages in measured performance and energy efficiency, especially in the performance of vision, voice and other scenarios. In addition, the company has also upgraded and optimized the basic system software platform, which effectively improves the development and deployment efficiency and reduces the learning cost, development cost and migration cost of users. At the same time, the company's talent team continues to improve and develop, and the combat effectiveness of the company's team has been significantly improved. At the end of 2021, the company has a total of 1213 R & D personnel, accounting for 81.03% of the total number of employees in the company; 77.00% of the R & D personnel have master's degrees or above. The company's R & D team has a reasonable structure and comprehensive skills, which strongly supports the company's technological innovation and product research and development.
Cambrian-U main business: applied to all kinds of cloud servers, edge computing equipment, terminal equipment in the development, design and sales of artificial intelligence core chips, as well as to provide customers with a wealth of chip products and system software solutions
Cambrian's 2022 first quarter report shows that the company's main revenue was 62.9907 million yuan, up 74.34% year-on-year; net profit attributable to the mother -287.3835 million yuan, down 39.78% year-on-year; deduction of non-net profit - 358.8841 million yuan, down 60.85% year-on-year; of which in the first quarter of 2022, the company's single-quarter main revenue was 62.9907 million yuan, up 74.34% year-on-year. Single-quarter attributable net profit - 287.3835 million yuan, down 39.78% year-on-year; single-quarter non-net profit - 358.8841 million yuan, down 60.85% year-on-year; debt ratio of 12.54%, investment income of 15.1603 million yuan, financial expenses - 15.5834 million yuan, gross profit margin of 56.58%.
A total of 4 institutions have given ratings, 2 buy ratings and 2 overweight ratings in the last 90 days; the average target price of institutions in the past 90 days is 83.48.
Here is detailed profit forecast information:
Margin data shows that the stock's net outflow in the past three months was 68.3508 million, and the balance of financing decreased; the net outflow of margin was 71.6097 million, and the balance of margin decreased. The Securities Star Valuation Analysis Tool shows that the Cambrian (688256) good company rating is 2.5 stars, good price rating is 1 star, and valuation comprehensive rating is 2 stars. (Rating Range: 1 ~ 5 stars, maximum 5 stars)
The above content is compiled by Securities Star based on public information, if you have any questions, please contact us.
![Apple M1 Max MacBook Pro ProRes benchmark[fig]](data:image/gif;base64,R0lGODlhAQABAIAAAP///wAAACwAAAAAAQABAAACAkQBADs=)