NEWS  /  Analysis

The Development of AI Chips Should Not Rely Solely on Advanced Processes, Says Huawei Executive

By   xinyue  Jul 10, 2024, 5:52 a.m. ET
Threads (10)

This makes Huawei the top player in China's AI computing power integration service market.

AsianFin--The development of AI chips should not solely rely on the most advanced process technology, said Zhang Ping'an, the executive director and CEO of Huawei Cloud Computing Technologies, at the recently-concluded 2024 World Artificial Intelligence Conference (WAIC).

He emphasized that structural innovation to build a sustainable AI computing power infrastructure and technology in China is crucial.

"AI development relies on data, algorithms, and computing power. China must develop our AI under the assumption that our current computing power is limited, said Zhang.

"I believe we need to consider collaborative innovation in the integration of cloud, network, edge, and core architectures for computing infrastructure. Through this collaborative innovation, we can build a solid foundation for AI computing power," Zhang explained, adding "We should not blindly pursue expectations for edge chips and advanced processes as the sole basis of AI infrastructure. If we rely solely on whether we have the most advanced process AI chips, we will not lead in AI if we don’t have that. This is wrong. Instead, we need to focus on how to shift the demand for chips and edge AI computing power to the cloud. In the cloud, we can establish China's advantages and build our network bandwidth effectively."

The overall scale of China's intelligent computing service market, driven primarily by AI-specific computing power such as GPUs, FPGAs, and ASICs, reached 11.41 billion yuan (US$ 1.57 billion) in the second half of last year, marking an 85.8% year-on-year increase, according to a report released by global research firm IDC.

The report highlights that the intelligent computing integration service market exhibits a "one super power and multi-great power" landscape, with Huawei leading the market thanks to its advanced chip capabilities and comprehensive service offerings. H3C, Baidu, Cambricon, and China Electronics Cloud are among the top five companies.

This makes Huawei the top player in China's AI computing power integration service market.

Zhang elaborated on Huawei's innovative direction of transferring AI computing power demands from the edge to the cloud through fiber optic and wireless networks. By doing so, Huawei aims to achieve seamless AI computing power collaboration between the edge and cloud, thus maintaining rich functionalities at the edge while significantly reducing power consumption and dependence on chips.

In terms of network technology innovation, Huawei has developed the 5G-Advanced network, which is optimized for AI. This network offers a 10-fold increase in both uplink and downlink bandwidth compared to 5G, and latency is reduced from 10 milliseconds to 1 millisecond, providing a solid network guarantee for edge computing power transitioning to the cloud and edge-cloud computing power collaboration.

China's abundant fiber optic networks and wireless 5G-A networks will establish China's AI computing power network, creating a leading edge for China.

For cloud infrastructure, Huawei Cloud has introduced the new CloudMatrix architecture, which aligns with the demands for ultra-large-scale computing power in terms of scale, expansion mode, and usage mode. Moreover, leveraging cloud technology, Huawei Cloud has optimized the Ascend cluster end-to-end, achieving uninterrupted training of trillion-parameter models for 40 days with an average fault recovery time of less than ten minutes, demonstrating significant advantages in long-term stability and fault recovery.

At the WAIC exhibition, Huawei Ascend disclosed that the Ascend training cluster supports 100% of mainstream large models, reducing the development cycle of state-of-the-art models from months to weeks. Basic operator support reaches 100%, and MindSpore enhances dynamic-static unification and multi-level compilation optimization efficiency by 200%. For large model inference, Huawei meets latency constraints of <50ms, achieving up to a 6-fold increase in throughput performance.

Zhang emphasized that China's AI development should aim to establish a globally leading position in large models within various industries. If all sectors actively embrace AI and open up industry business scenarios, China has the potential to lead globally in the B2B sector.

Currently, Huawei Cloud is advancing the iteration and implementation of large model technology. In June, the Pangu large model 5.0 was officially released, offering richer innovative applications and practical implementations in fields such as autonomous driving, industrial design, architectural design, embodied intelligence, digital content production, high-speed rail, steel, meteorology, and pharmaceuticals.

In the steel industry, the Pangu large model has already been deployed on a hot rolling production line at Baosteel, improving the accuracy of steel plate predictions by 5%, potentially increasing annual steel plate production by 20,000 tons and generating an additional revenue of over 90 million yuan, Zhang noted.

Huawei's Ascend ecosystem and large model innovations have been implemented in over 30 industries and 400 business scenarios. Since the launch of generative AI cloud services, Huawei Cloud has served more than 600 pioneering enterprises, Zhang added.

Through full-stack innovation in AI hardware products, foundational software, cloud services, and Huawei's global data centers and infrastructure, Huawei is accelerating the deep adaptation of Huawei Ascend to various industries, becoming a crucial part of China's AI development, he said.

"As the era of the AI wave arrives, let us build China's AI industry together, seize opportunities, and usher in a new future for AI," Zhang remarked.

Related threads