Shulou(Shulou.com)11/24 Report--

On September 13-14, the 2023 Open data Center Conference was successfully held at the Beijing International Convention Center. With the theme of "Computing enables unlimited openness", the conference focused on the technological trend of the computing industry and the ecological construction of the data center industry. Chen Bochao, head of the switch product line product management department of Xinhua third Group, was invited to attend the network forum and delivered a keynote speech, sharing the forward-looking thinking and practice of how Xinhua three used intelligent networks to support the large-scale and complex computing needs of data centers while achieving a breakthrough in speed and efficiency in the AIGC era.

At present, the application of AIGC is developing at an astonishing speed. Data show that from 2018 to 2022, the number of large model parameters soared from 100 million to 540 billion, and the demand for computing power also exploded. However, due to the limitations of Moore's Law, the scale of fat tree architecture and the efficiency of parallel communication, network data transmission has become a bottleneck to improve computing power. In order to break the network shackles of computing power evolution and help Baiye enjoy the AIGC technology dividend, Xinhua three groups launched a new generation of intelligent computing network solutions, focusing on the four key directions of ultra-wide network, extremely low latency, end-to-end network storage integration, and low-carbon energy saving, to fully meet the key technical requirements and application experience needs of the network in the era of intelligent computing.

Ultra-wide network: accelerating the landing of high-performance network technology

In AI cluster, high-performance GPU, trillion-parameter training scale and high-speed parallel computing requirements require higher network bandwidth to support efficient communication between computing nodes. As a leader in advanced network technology, Xinhua 3 Group continues to lead the production process of high-speed network technology, from the completion of the industry's first single-frame 768 to 100G performance test in 2017, to the industry's first 800G CPO silicon data center switch in 2023, and verified by professional institutions, Xinhua 3 has created a full range of intelligent switch products and solutions, leading the evolution of the ultra-wide intelligent network.

On the basis of high-speed network, in order to maximize bandwidth benefits, Xinhua Group 3 integrates intelligent lossless network features into its products to solve the problem of cluster performance loss caused by Ethernet packet loss. Among them, the global load balancing technology can identify the optimal path according to the node flow congestion state of the whole network switch and the whole network topology, increase the overall network throughput to 90%, and greatly shorten the AI training cycle. SprayLink and AgileBuffer functions bring solutions to the elephant flow and mouse flow characteristics of the AIGC scenario, which can improve the link bandwidth utilization by 95% and achieve zero packet loss for high-priority services.

Extremely low latency: optimizing the operational efficiency of intelligent computing networks

In the large-scale computing environment, the delay jitter will also affect the AI training efficiency. How to reduce the computing communication delay is another core problem for the intelligent computing center to fully release the computing power. Xinhua Group uses the leading design to reduce the transmission delay of the equipment. The H3C S9827 series integrates CPO and LPO silicon technology, and the delay can be reduced by up to 25% compared with the traditional design. In addition, in order to improve the overall availability of the intelligent computing network, Xinhua three launched the H3C S12500AI series based on DDC ultra-low latency architecture, which can effectively reduce end-to-end latency and jitter.

Chen Bochao said that under the traditional fat tree architecture, large-scale computing requires a three-tier network architecture with large forwarding delay, which is easy to cause unnecessary multi-hop forwarding and lack of load balancing optimization technology for multi-path. In order to further optimize the operational efficiency of the intelligent computing network, Xinhua Group 3 implemented the Spine-Leaf two-tier architecture based on ultra-wide products, optimized the transmission efficiency of the whole process, increased the bandwidth utilization of the entire network by 20%, and reduced the end-to-end delay by 15%.

End-to-end network storage integration: provide full-stack products for computing power construction

At the same time, with the growth of AI cluster size, the difficulty of deployment of computing, storage, network and other infrastructure also increases. Xinhua Group 3 integrates the advantages of the "cloud-network-computing-storage-end" industry chain to establish the "end-to-end network storage integration" capability around the AI model, combined with powerful AI servers, lossless network solutions, and a full range of high-performance storage products, from models to infrastructure one-stop application deployment to form end-to-end smart computing solutions to meet the omni-directional needs of AI computing power construction.

Low-carbon energy saving: multi-dimensional breakthrough of power consumption dilemma

High computing power is often accompanied by high energy consumption. How to build an efficient and energy-saving green data center is an important proposition related to the sustainable development of AIGC. Xinhua Group breaks through the dilemma of power consumption in data centers from multiple dimensions, by increasing research on optical technology to increase network speed while reducing power consumption and heating; by tapping the energy-saving potential of equipment to create efficient power supply schemes and low-power fans to improve energy efficiency; through the development of liquid cooling technology, create cold plate, immersion liquid cooling and other full-stack products, start from the bottom to build a green digital infrastructure.

With the vigorous development of the intelligent computing industry, network technology is bound to usher in more major innovations. Xinhua three groups put forward four major development trends of network technology based on their deep insight into the needs of the industry and the direction of technological evolution. Including the accelerated listing of ultra-large bandwidth products, the emergence of Al-oriented Ethernet protocol framework, the continuous penetration of heterogeneous computing resource exchange technology represented by CXL, and the gradual popularization of embedded network operation and maintenance model, these four directions will also be the key R & D investment and technology development direction of Xinhua 3.

As a leader in the network field, facing the future, Xinhua Group will take the industry trend as the guide, uphold the concept of "intensive cultivation and pragmatism, and endow the times with wisdom", continue to promote the exploration and practice of network technology under the AIGC scene, constantly refine network products and solutions, and promote Baixing and Baiye to move forward towards comprehensive intelligence.

Welcome to subscribe "Shulou Technology Information " to get latest news, interesting things and hot topics in the IT industry, and controls the hottest and latest Internet news, technology news and IT industry trends.