With the Surge in AI Demand, Is China's Computing Power Ready?

[baoshi.rao]

Benefiting from the latest wave of AI enthusiasm, the stock price of U.S. GPU chip manufacturer Nvidia has tripled in less than eight months. By the end of May, the company's market capitalization surpassed $1 trillion, successfully joining the "trillion-dollar club."

Much of the credit likely goes to OpenAI's launch of ChatGPT, a major industry milestone. Since its debut in November 2022, ChatGPT's registered user count skyrocketed to 100 million in just three months.

These two sets of data unmistakably reveal the unprecedented demand for computing power driven by AI. This has also put pressure on Chinese scholars, experts, and entrepreneurs in the computing industry—facing a challenging external environment, how should China's AI computing power services respond?

On May 29, the China Intelligent Computing Industry Alliance initiated the 2023 China Computing Power Development Symposium, listing the above question as a key agenda item and inviting experts to share their insights.

Supercomputing Internet "Assists" AI Computing Power Services

"To develop large models, you must have massive computing power. As the saying goes, 'Great power brings miracles'—the greater the computing power, the more likely it is to achieve breakthroughs," said Zheng Weimin, an academician of the Chinese Academy of Engineering and a professor at Tsinghua University's Department of Computer Science. He noted that ChatGPT is a product built on large-scale computing infrastructure and questioned where China could find sufficiently powerful AI computing resources to train even more advanced large models.

"This requires new types of computing infrastructure specifically designed for such services," he added. "Of course, there are many issues to address, including network connectivity, data transmission, and more."

The answer may be close at hand. In April this year, a major event occurred in China's computing sector: the Ministry of Science and Technology launched the National Supercomputing Internet initiative. According to the plan, this initiative will connect numerous supercomputing centers across the country via a computing power network, applying internet-based operational models to supercomputing centers. The goal is to create an integrated computing power service platform, enabling coordinated scheduling of computing resources, lowering the barriers to supercomputing applications, and driving computing technology to higher levels.

As previously explained by Cao Zhennan, Deputy Director of the National Research Center for High-Performance Computing Engineering Technology, the National Supercomputing Internet platform is currently in the development phase. It aims to establish an advanced, innovative, high-quality, and ecologically sound overall framework.

How much will the Supercomputing Internet platform contribute to AI computing?

"Although computing power is often categorized into supercomputing and AI computing, the two share many similarities in models and applications, differing mainly in their precision requirements," Cao noted. Currently, the demand for AI computing power far exceeds supply, and performance often takes precedence over precision. However, if computing power becomes more accessible and affordable, "some AI applications may start demanding higher precision."

Cao believes the boundary between supercomputing and AI computing is becoming increasingly blurred. At present, AI computing power is unevenly distributed across regions and industries, creating significant gaps in many sectors. One of the goals of the Supercomputing Internet platform is to address this imbalance, enabling interconnected and shared computing resources.

At the symposium, Qian Depei, an academician of the Chinese Academy of Sciences and head of the Supercomputing Internet Expert Group, also shared his profound insights on computing services.

"Our computing power infrastructure should aim to provide not just 'foundational stones' but the actual software or services users need. This is the essence of the computing power internet—only then will people stop caring about which processors deliver the computing power," said Qian Depei. He noted this would impact the operational and service models of supercomputing and AI computing centers, ultimately integrating computing power into all sectors as part of daily production and life. "Of course, there's still a long way to go."

Computing Power Services Require Rational Planning to Avoid Waste

Regardless, as AI technology and applications evolve, the demand for computing power is growing inexorably.

"We anticipate that generative AI, represented by ChatGPT, will increasingly dominate data center computing," said Chen Wenguang, a professor at Tsinghua University. He observed that AI computing centers show higher utilization rates than traditional supercomputing centers, "indicating AI's deepening penetration across economic and social sectors."

"We should embrace this trend," said Lu Zhonghua from the Chinese Academy of Sciences. While opinions vary on computing infrastructure and interconnectivity, one consensus remains: the surging demand for computing power, especially for AI applications.

Lu proposed three recommendations:

1. Maximize existing supercomputing/AI computing resources over the next 5-8 years to avoid waste;
2. Embrace domestic AI chips to build China's ecosystem for large model development;
3. Market forces should drive the construction of integrated AI computing platforms rather than relying solely on top-down planning.

However, AI R&D users still expect national-level planning for computing power services. Academician Chen Runsheng emphasized the need for strategic division of labor: "China needs OpenAI-level models but also diverse computing services tailored to different capabilities—avoiding redundant efforts that waste resources."

Building a Domestic Computing Power Network Ecosystem is Crucial

Zhang Yunquan highlighted the lack of unified standards for computing power interconnectivity since its conceptualization, urging national-level roadmaps to prevent missteps and resource waste.

The National Supercomputing Internet Platform has taken initial steps in this direction. Cao Zhennan identified application development as the key bottleneck, advocating for coordinated hardware-software progress and domestic ecosystem growth in computing networks.

"A crucial objective in building the supercomputing internet is to establish China's supercomputing and computing industry ecosystem. We hope different entities can form standards through this ecosystem, thereby creating more reasonable market models," said Cao Zhennan. This initiative also represents an opportunity for China's computing service market to enter a new development phase amid growing computational demands from AI applications.