CEO of Huawei Cloud: the notion that China cannot lead in AI without advanced AI chips must be discarded
At the World AI Conference in Shanghai (WAIC), I heard some relatively positive and optimistic views on China's position in future global AI competition. For example, Wang Jian(王坚), an academician of the Chinese Academy of Engineering and founder of Alibaba Cloud, highlighted China's advantages in energy and the benefits they bring to AI research and model training. He pointed out that the bottleneck for AI development in the United States is electricity, which is not a problem for China. China's annual electricity generation is more than the combined total of the United States, Japan, and Russia.
Similarly, Mr. Zhang Ping'an(张平安), Huawei's Executive Director and CEO of Huawei Cloud, emphasized China's advantages in bandwidth, network, and energy. He proposed leveraging these advantages to address AI computing power issues in the cloud, considering this to be a direction China can firmly rely on.
Zhang emphasized that China should not rely solely on the most advanced AI chips but should leverage its advantages in bandwidth, network, and energy to address AI computing power issues through architectural innovation in the cloud. He called for rapidly opening industry scenarios to accelerate AI applications across various sectors and build a globally leading AI ecosystem. Additionally, Zhang introduced the successful applications of Huawei's Pangu Model 5.0 in multiple industries, demonstrating Huawei's leadership in advancing AI technology and industry innovation.
His key points include:
China should achieve collaborative innovation through architectural innovation, integrating chips, end devices, cloud, and networks to build a sustainable AI computing infrastructure.
Due to the limitations of end devices in chip technology, power consumption, and size, China should offload more AI computing power demands from edge devices to the cloud, leveraging cloud computing to enhance AI application performance.
China needs to improve network bandwidth further and reduce transmission latency to realise edge-cloud collaborative computing. Utilizing the 5G-A network and abundant fibre optic resources will support efficient data upload from the edge to the cloud.
Build high-efficiency data center infrastructure through the innovative Cloud Matrix architecture, achieving pooling and equal management of computing resources to meet future large-scale AI computing demands.
Utilize China's advantages in green power and other energy sources, innovating in cooling architectures to reduce data center energy consumption and build a green, efficient computing foundation.
Accelerate the opening of business scenarios across industries, promote rapid AI development in B-end applications, and leverage China's abundant industry scenarios and software innovation community to build a globally leading AI application ecosystem.
Below is the transcript of the speech by Zhang Ping’an:
First of all, I would like to thank the organizers for inviting me and allowing me to share my views on how we can build a sustainable AI computing infrastructure and AI computing technology in China through architectural innovation.
Artificial intelligence is profoundly impacting our lives and work. We are rapidly transitioning from digitalization to intelligence, and the development of AI is reshaping numerous industries. AI development relies on data, algorithms, and computing power. In China, we must develop our AI, assuming that our computing power is limited.
So, what path should China take for AI innovation? When it comes to computing power infrastructure, we need to consider the integration of chips, edge, cloud, and networks. Through this synergy, we must collaborate on architectural innovation to build our AI computing infrastructure.
Firstly, we cannot blindly pursue expectations for the process technology of edge-side chips. We cannot base the construction of our AI infrastructure solely on whether we have the most advanced process AI chips. The notion that without these advanced chips, we cannot lead in AI must be discarded in China.
We need to think more about how to shift the demand for AI computing power from chips and edge devices to the cloud because, in the cloud, we can build on China's advantages.
The era of intelligence is rapidly approaching. In the future, all connected end devices will be intelligent devices. These connected end devices will have high demands for AI computing power. We cannot rely solely on the end devices to meet the AI computing power requirements because they are limited by chip technology, power consumption, and size, especially in mobile phones, PCs, and various industrial intelligent end devices. We want industrial intelligent end devices to last for one or two years without replacement and minimise power consumption's impact.
We cannot solely provide too much AI computing power on the end devices’ side. For example, running a large model with 700 million parameters on a mobile phone forces us to consider reducing it to 300 million or even 150 million parameters. Running a billion-parameter model offline on the end devices side is unfeasible!
Given the limitations of edge computing power, we should consider offloading it to the cloud. When offloading to the cloud, we need to consider how to transmit data to the cloud and whether it can be computed there.
We have already initiated many innovations in this area, such as cloud-based offices, cloud photography, cloud phones, cloud gaming, and cloud design. We have dedicated project teams exploring how to offload edge computing demands to the cloud. By leveraging the cloud's rich computing power, we can maintain the rich functionality of edge devices while significantly reducing power consumption and reliance on chips.
People enjoy using Huawei phones for photography. Huawei phones have different lenses, and every time we press the shutter, the phone captures at least six images, each between 8MB and 12MB in size. While our phone photos look clear, the original images are even clearer. We upload these six original images to the cloud for algorithmic processing when edge computing power is limited. The difference between the photos can be substantial, even approaching the quality of a DSLR, with clearer and more detailed images. Therefore, we achieve better results by offloading computing power to the cloud.
For us, the innovation direction is to avoid making the edge too complex. Why not offload edge computing power to the cloud? Utilizing the extensive computing power of the cloud allows AI to provide greater value.
Secondly, as discussed earlier, seamless collaboration between edge and cloud computing power requires a robust network.
We must construct a network with sufficient bandwidth to build China’s advantages in the cloud. Our current networks are strong in downstream capabilities but lack in upstream. Therefore, we must focus on building large bandwidths with low latency for both upstream and downstream. This is our direction for network innovation.
China Telecom, Mobile, and Unicom have extensive fibre optic networks. Coupled with the speed of our fibre optics and abundant cable resources, this is a unique advantage for China. We should leverage this advantage for innovation.
China leads the world in 5G network construction, and last year, we began innovating 5G-A networks. The 5G-A network offers ten times the bandwidth of current 5G networks, with upstream and downstream speeds ten times faster. More crucially, it reduces transmission latency from 10 milliseconds to 1 millisecond.
With the extensive 5G-A wireless network and abundant fibre optic transmission bandwidth, we can offload edge data to the cloud, addressing edge computing power demands in the cloud.
We can establish a robust network for the AI era with these solid network foundations and China's unique advantages.
Thirdly, to build infrastructure in the cloud, we need to construct a sustainable AI computing base through data center space, bandwidth, and energy. Innovation in data centers allows us to supply computing power in the cloud efficiently. We have made significant innovations in cloud data center infrastructure.
With the advent of AI computing, we realized the need to move away from the CPU-centric computing architecture. Previously, CPU servers were CPU-centric, with all external devices considered peripherals. Regardless of their function, all data had to be processed by the CPU. Given the AI focus, why should traditional data centers remain compute-centric?
We consider treating all computing resources and infrastructure in data centers equally. This is the foundation of our newly released Cloud Matrix architecture, where all resources in the data center can be pooled and treated equally in the network. Based on this, we can design fully composable data centers. This allows us to match future large-scale computing needs with scalable and efficient resource usage.
Our cloud team, particularly with the Ascend AI cluster, has broken down barriers between CPUs, NPUs, DPUs, storage, and networks, optimizing them end-to-end. Each component no longer works independently; instead, they collaborate to complete AI computing tasks.
As a result, we can achieve 40-day uninterrupted training of trillion-parameter models in the cloud, six to ten times more efficient than offline. Every so often during training, we synchronize all parameters to prevent failures. Offline, parameter synchronization takes one to two hours, but in the cloud, it can be done in 10 minutes. Our team is exploring achieving one-minute synchronization, making cloud computing even more efficient.
Additionally, China has advantages in green power, hydropower, and other energy sources. As Academician Wang Jian mentioned, China’s electricity generation equals that of all G7 countries combined, with further significant energy investments.
Through innovative cooling architectures, we have significantly reduced the PUE (Power Usage Effectiveness) of data centers. Our newly built data center in Wuhu, using liquid and natural cooling, achieved a PUE of 1.1, making it the most energy-efficient data center in East China despite an average annual temperature 10 degrees higher than Guiyang.
Leveraging bandwidth, network, and energy advantages to address computing power issues in the cloud is a reliable direction for China. By building on these advantages, we can establish robust cloud data centers to provide a solid AI computing foundation.
Lastly, while China has many foundational models, we lack innovation in C-end and B-end applications. Although China excels in C-end internet innovation, we need to accelerate the opening of industry scenarios to rapidly build a global leading position in AI and large-scale models for industry applications.
China has thousands of industry scenarios and the world’s largest software innovation community. If all industries actively embrace AI and open up industry business scenarios, China has a significant opportunity to establish a leading global AI position in the B2B field.