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

[Shanghai, China, September 20, 2023] during Huawei full connection Conference 2023, at the Xinghe AI Network theme Summit of Datong Huawei, together with China Institute of Information and Communications and iFLYTEK Research Institute, has jointly released the "White Paper on Galaxy AI Network" (hereinafter referred to as the "white paper") to the world, which demonstrates the technical leadership of Xinghe AI Network in the artificial intelligence industry from three aspects: the development trend of AI business, network architecture and key technological innovation. The purpose of this paper is to promote the numerical intelligence transformation of the industry, promote industrial upgrading and cooperation, and provide a reference for the construction of a high-performance training network for the AI model.

Guo Liang, Chief engineer of Cloud Computing and big data Research Institute, China Institute of Information and Communication (left)

Zhao Zhipeng, Vice President of Huawei data Communication Product Line (middle)

Wang Jinyi, Vice President of iFLYTEK AI Engineering Institute (right)

The white paper points out that in recent years, the AI algorithm has entered the era of trillions of parameter models, and the computing power demand has increased nearly 100, 000 times. AI large-scale computing requires efficient cooperation with tens of thousands of AI processors, and needs to continuously optimize the network to improve parallel computing efficiency; and because of the high cost of AI processors, there is an urgent need to build a high-performance network with zero packet loss and high throughput, so as to fully release the efficiency of AI processors. At the same time, the training period of AI large model is long, and the traffic scale of Vanka cluster reaches 100,000, so effective operation and maintenance means are needed to reduce the average time between failures.

Based on the above challenges, the Galaxy AI network solution proposes the following innovations:

High performance: using NSLB (AI Accelerator) technology to increase network throughput to 98% and improve the efficiency of training by 20%

High reliability: sub-millisecond handover of faulty links using DPFR (Service sensorless self-healing) technology to ensure that the fault application is not aware of it.

Operable maintenance: provide visual operation and maintenance scheme, collect data with high precision, realize network one-button fault diagnosis, and improve the troubleshooting efficiency by 90% during training.

Large-scale: support Vanka cluster training, the calculation scale is 4 times that of the industry

Openness: provides super-integrated ethernet solution, which can be completely reused for ethernet ecology. Compared with the industry, the performance is basically the same, and the operation and maintenance cost is reduced by 30%.

At present, the development of AI model has changed from thousands of small models to a large model era. Xinghe AI network solution aims at the new requirements of the network, continuously optimizes the network architecture, innovates the network technology, provides a reference for building a high-performance training network oriented to the AI model, and promotes the development and progress of artificial intelligence technology.

At present, Xinghe AI network has been deployed commercially in 100 + enterprises around the world. Facing the future, Huawei hopes to work with many partners to promote the development of technology and the expansion of application scenarios, so as to achieve the sustainable development of artificial intelligence technology and the common prosperity of society.

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.