AI infrastructure company "Jiliu Technology" has successively completed Pre-A+ and Series A rounds of financing, with a cumulative total financing amount of over 100 million yuan. | This is an exclusive report from 36Kr.

Written by Wang Fangyu

Edited by Su Jianxun

36Kr has learned that "Jiliu Technology", a computing power network and computing power construction and maintenance service provider affiliated with Tsinghua University, has recently completed Pre A+ and Series A financing. The financing is jointly invested by China Merchants Venture Capital, Huatai Innovation, Xinglian Capital, and Guofang Innovation, with existing shareholders Zhuoyuan Asia and Lightspeed China Partners following up. The raised funds will be mainly used for product research and development, market promotion, and daily operations.

The historical investors of Jiliu Technology include multiple strategic investors and state-owned capital funds, with a cumulative total financing amount of over 100 million yuan.

Jiliu Technology was established in February 2023 and originated from the Network Security Laboratory of Tsinghua University. The company has overcome the distributed computing and communication challenges of AI infrastructure and has formed a series of key technologies in high-speed networking, collective communication, parallel framework, management and control scheduling, etc.

Since its establishment, Jiliu Technology has achieved an upgrade from hundreds of cards, thousands of cards to tens of thousands of cards, and from cluster networking, optimization to cluster operation and maintenance.

Up to now, the company has supported the landing of more than ten intelligent computing clusters, and the cumulative FP16 computing power of the constructed and optimized clusters has exceeded 40 EFLOPs. The company has served multiple leading users including Zhipu AI, SenseTime, data centers, operators, and local state-owned enterprises.

As one of the three major elements promoting the development of artificial intelligence, computing power is the "engine" and core driving force of artificial intelligence, and has also become a new type of infrastructure following water, electricity, gas, roads, and networks.

The "China Comprehensive Computing Power Index Report (2024)" released in September this year points out that in the past 20 years, the demand for intelligent computing power in China has increased by hundreds of billions of times. With the rapid development of industries such as AI and embodied intelligence, the scale of intelligent computing in China will further increase rapidly. There will also be an increasing number of "tens of thousands of card clusters" composed of 10,000 or more accelerator cards.

The larger the computing power cluster, the higher the probability of network communication and GPU component failures. Quickly building a large-scale and stable computing power cluster while achieving high cost performance is a global challenge. And this is exactly the goal that Jiliu Technology wants to achieve.

Just as the Internet giant Cisco mainly solves the problems of Internet expansion and connection in the Internet era, the role of computing power network service providers is equivalent to that of Cisco in the era of large models. Computing power networks have important strategic significance. Recently, Nvidia has been subject to an anti-monopoly investigation by the State Administration for Market Regulation due to its acquisition of the high-performance network company Mellanox. The dedicated network interconnection equipment and high-speed Ethernet network cards for GPUs need to break the blockade.

"The demand for computing power in artificial intelligence is very different from that in traditional IT. Therefore, the network architecture needs to be redesigned. Relevant technologies are still in the initial stage, and there is still a lot of imagination space in distributed computing frameworks, hardware devices, and even chips." Hu Xiaoe, the founder of Jiliu Technology, told 36Kr.

At present, there are very few teams in China with the technical ability to build and maintain large-scale clusters of more than 10,000 cards, and the industry generally lacks the experience and standards for the landing of computing power clusters. And Jiliu Technology is one of the very few AI infrastructure manufacturers in China with the experience of landing 10,000-card clusters, and has experienced the upgrade of the domestic computing power scale from thousands of cards to 10,000 cards.

In terms of the team, the founder Hu Xiaoe is a Ph.D. and postdoctoral fellow from Tsinghua University, and a visiting scholar at the University of California, Berkeley. He has been conducting research in the direction of high-performance networks and high-performance computing for more than ten years. Before starting his business, he has already landed the country's first operator-level Tbps programmable network product. Team members come from first-class universities such as Tsinghua University, Peking University, Beijing University of Posts and Telecommunications, and Beihang University, as well as Internet and equipment manufacturers such as Alibaba, Meituan, and ZTE. Many of the members have more than twenty years of production and research experience.

Benefiting from the explosive growth of the track, Jiliu Technology is also growing rapidly. It is understood that the company confirmed an income of several tens of millions of yuan in 2023, and the expected income in 2024 will exceed 100 million yuan.

In terms of products, Jiliu Technology currently has three main parts: the computing power scheduling and optimization platform, the computing power construction and operation and maintenance platform, and the high-speed interconnection hardware. In addition to the entire set of computing power cluster construction solutions, the company has productized and gradually implemented products at the three levels of cluster management, computing framework, and high-speed network, and strives to improve the delivery efficiency, GPU utilization rate, and cluster stability as much as possible.

Hu Xiaoe told 36Kr that the computing power cluster solution of Jiliu Technology can improve the performance of the GPU cluster by more than 10% in the production environment, helping customers save tens of millions of yuan in the landing of thousands of cards and hundreds of millions of yuan in the landing of tens of thousands of cards.

For the construction of a larger-scale 100,000-card cluster, Hu Xiaoe said "No one in China knows how to build and use a 100,000-card cluster yet, and further exploration is needed." However, the Jiliu Technology team has already started to invest in the project research and development of 100,000-card simulation and future architecture design.

In addition, the localization of high-speed computing power interconnection hardware is also underway.

Hu Xiaoe said that the core components of the computing power network, such as network cards, cables, and switches, are expected to achieve end-to-end localization by 2025. Through the method of combined software and hardware design, the algorithm, process, and supply chain will be optimized to drive the development of chips with the system, achieving an open hardware ecosystem and independent and controllable software.

"With the rapid development of the artificial intelligence industry, a 100,000-card cluster will definitely be generated in China in the next one to two years. At the strategic level, this is a process of matching high-quality supply and demand." Hu Xiaoe said that Jiliu Technology will continue to exert efforts in the networking, optimization, and maintenance of ultra-large-scale clusters to contribute to serving the national grand strategy.