Research Note

Abstract

LightCounting reports from China Cloud Network Conference

At the China Cloud Network Conference over 700 attendees gathered at the JW Marriott hotel in central Beijing on April 10-11, 2024, to discuss and explore how AI could help communications service providers (CSPs) move towards autonomous networks. There is a lot of progress, but Level 4 autonomous networks remain at least a year away, even for China Mobile – the largest service provider in the world.

Ao Li, Chief Engineer at the China Academy of Info & Communication Technology (CAICT), stated that currently the three main CSPs in China – China Mobile, China Telecom and China Unicom – have all reached TM Forum’s Level 3 classification for autonomous networks.

Service providers in China have already deployed many compute nodes at the network edge to support new applications. The top priority for 2024 is AI Clusters at the edge for training and inference and very large (up to 80,000 GPUs) clusters for training. The training clusters are known as intelligence computing centers. China Mobile is expected to play a leading role in these deployments with 50 intelligence computing centers planned to be up and running by the end of 2025.

Tencent disclosed even more ambitious designs of AI Clusters (up to 128,000 GPUs), supported by internally designed 51.2T Ethernet switches. H3C discussed their latest AI switch, S12500AI, which could support up to 32,000 GPUs. More importantly, it employs a distributed disaggregated chassis (DDC) architecture, separating line cards and fabric cards into separate chassis. H3C built an internal network to test and verify a large model, it consists of 16 units of 128 GPU cards, with different GPU systems in them from NVIDIA (Mellanox), Broadcom and TS.

Wei Leping, Chairman of the Science & Technology Committee, China Telecom (Group), gave an address at the event. He pointed out that the biggest difficulties in AI development in China are access to GPUs and the application ecosystem around them. Nvidia has the leading share of the deployed GPU systems in AI and high-performance computing clusters in China, whether at telecom service providers or Cloud companies.

The US government’s ban on selling high-end GPUs to China, in place since 2022, has forced Chinese vendors to develop their own solutions, backed financially by the Chinese government. At present, Huawei is taking a lead position domestically for GPU system development including the related environment for software development. The Ascend 910B, Huawei’s latest GPU released in 2023, will be in volume shipment this year. It is similar in performance to Nvidia’s A100.

Full text of the research note is available to LightCounting clients at: https://www.lightcounting.com/login