Recently, the International Forum on High Performance Computing and Artificial Intelligence Collaborative Innovation (HACI) was held in Shenzhen, Guangdong. Shao Guanglu, Director of the Science and Technology Committee of China Telecom, was invited to attend the meeting and shared China Telecom's practices and thoughts on intelligent computing clusters and networks.
"Computing power is an important driving force for the development of artificial intelligence, affecting the speed of iteration and innovation of large models, and also affecting the speed of economic development." Shao Guanglu introduced that in terms of intelligent computing, China Telecom has focused on promoting two tasks. The first is to build a large-scale intelligent computing cluster. At present, two Wanka clusters in Shanghai and Beijing have been put into production. Intelligent computing clusters for AI large model training need to be upgraded from IDC to AIDC, focusing on six key indicators such as model computing power utilization (MFU), linear acceleration ratio, stability, latency, packet loss rate, and GPU utilization. The design, construction, operation, and optimization of intelligent computing clusters are continuously carried out to improve the efficiency and effectiveness of large model training. The second is to use intelligent computing networks and computing power scheduling platforms to achieve the interconnection and interoperability of computing power resources across the country, and call them on demand. The self-developed "Xiran" computing power scheduling platform has achieved a computing power scheduling of 21EFlops across the country. China Telecom actively explores the integration technology of "cloud, network, and intelligence", and is committed to building an intelligent computing network with elastic ultra-wide, high-speed and lossless, and "cloud, edge, and end" full-domain scheduling to meet the differentiated needs of different enterprises and industries for large model training and reasoning.
Shao Guanglu said that China Telecom has upgraded from a telecom operator to a telecom + cloud computing operator. In the AI era, China Telecom will also upgrade to an intelligent computing service provider and a large model service provider, providing computing services and model services.
Shao Guanglu pointed out that the networking of computing power has become a development direction, and proposed: First, promote the open source and openness of the computing power ecosystem, provide a basic development environment for heterogeneous computing power interoperability, and connect various heterogeneous hardware with unified standards to facilitate the free switching of large models between heterogeneous computing power. Second, promote co-construction and sharing. Generative artificial intelligence is a major scientific and technological innovation and industrial transformation, which requires early layout and rapid advancement. It requires academia and industry to strengthen cooperation with a strategic vision, concentrate on breaking through the underlying technology, and promote industrial innovation and industrial ecosystem construction.