Huawei's Initiatives in AI Development in China
TapTechNews July 4th news, the 2024 World Artificial Intelligence Conference opened today at the Shanghai World Expo Center. Zhang Ping'an, the Managing Director of Huawei and CEO of Huawei Cloud, said that the development of AI in China cannot be separated from the innovation of computing power infrastructure, and it is necessary to dare to open up industry scenarios to make AI take the lead in industry applications.
Zhang Ping'an proposed that on the path of China's AI development, what should be pursued is to build a globally leading position of large models in the industrial field. If all industries actively embrace AI and actively open up their business scenarios, China has a great chance to build a global leading advantage in the 2B field. Take the steel industry as an example. Currently, Pangu large model has been launched on a hot-rolled production line of Baosteel, and the prediction accuracy of steel plates has increased by 5%, and it is expected to produce 20,000 more tons of steel plates every year, increasing revenue by more than 90 million yuan.
Zhang Ping'an said that Huawei is building a sustainable AI computing power foundation through coordinated innovation in the cloud-network-device-chip architecture, including three aspects: on-chip computing power going to the cloud, network architecture upgrade for AI, and cloud infrastructure system architecture innovation.
In terms of on-chip computing power going to the cloud, Huawei releases the AI computing power demand on the device side to the cloud through optical fibers and wireless networks, and obtains seamless AI computing power through end-cloud coordination. Through the computing power on the cloud side, the device side not only maintains rich functions but also greatly reduces power consumption and the dependence on chips.
In terms of network technology innovation, Huawei has 5G-Advanced (TapTechNews note: also known as 5.5G) network, and its uplink and downlink bandwidths have both increased by 10 times compared to 5G, and the latency has been reduced from 10 milliseconds to 1 millisecond, providing a solid network guarantee for on-chip computing power going to the cloud and end-cloud computing power coordination.
In terms of cloud infrastructure, Huawei Cloud has released a brand new CloudMatrix architecture, which matches the demands of super-large-scale computing power in terms of computing power scale, expansion mode, and usage mode. At the same time, Huawei Cloud conducts end-to-end optimization for the Ascend cluster. The training of a thousand-billion-parameter model on the cloud can achieve 40 days without interruption, with an average fault recovery time of less than 10 minutes.
