China Telecom completed the industry's first kilo-card distributed lossless intelligent computing network verification based on 800GWDM existing network

1.Currently, as the number of smart computing cards reaches 1,000 or even 10,000 cards, the required bandwidth is often as high as hundreds of T bits or even more than P bits. class. Therefore, the large bandwidth, high reliability and high efficiency of the optical transmission system are the key to ensuring high computational efficiency of distributed training.

2.China Telecom Research Institute, China Telecom Beijing Company, and Tianyi Cloud Company jointly provide large-scale distributed clusters with 1024 cards based on 800G C+L technology on China Telecom’s existing network. Capacity bandwidth enables distributed training of large models with 100 billion parameters over 120 kilometers, and the performance reaches more than 95% of centralized training.

3.In response to the problem of large bandwidth for data transmission, China Telecom adopts single-wavelength 800G technology with high-order modulation format to improve spectrum efficiency, and cooperates with the current industry hotspot C+L band technology To achieve ultra-large transmission bandwidth, Huawei transmission equipment was used between China Telecom Wuqing and Runze computer rooms, and a large-bandwidth interconnected intelligent computing verification network was built through multiple loopbacks, with the distance reaching 120km.

4.In response to the high reliability problem of data transmission, China Telecom has completed abnormal test experiments such as link errors, wavelength failures, fiber failures, etc. The results show that an 800G service wave interruption will This results in a reduction of more than 40% in computing efficiency, and optical fiber failures of more than 100 milliseconds can cause a significant reduction in computing efficiency or even interrupt training. According to official information, WSON re-route recovery technology is used to control the re-route recovery time between two points within 50ms, which can ensure high-reliability interconnection of distributed intelligent computing services and maximize computing efficiency.

5.In response to the problem of high efficiency of transmission links, China Telecom proposed a minute-level wavelength dynamic demolition and construction solution to achieve collaborative time-sharing multiplexing of computing and networks, effectively improving network resources. Utilization. This verification laid the foundation for highly reliable collaborative scheduling of computing power across regions, levels, and subjects.

The above is the detailed content of China Telecom completed the industry's first kilo-card distributed lossless intelligent computing network verification based on 800GWDM existing network. For more information, please follow other related articles on the PHP Chinese website!