China is intensifying efforts to establish a domestic graphics processing unit (GPU) market as it seeks to reduce its reliance on Nvidia chips, which have become scarce due to US export restrictions. The restrictions, initiated to prevent advanced technologies from bolstering China's military capabilities, have prompted significant changes in the country's approach to AI and computing.
The export sanctions, which specifically barred Nvidia's A100 and H100 GPUs from reaching China in August 2022, led Nvidia to launch modified alternatives, the A800 and H800. However, these too faced bans, further limiting China's access to essential technology. As a result, the Chinese government and various stakeholders now focus on developing homegrown GPU solutions, hoping to advance their AI capabilities despite the evident challenges.
Efforts by Chinese start-ups to innovate in GPU hardware and software have yielded notable progress in recent months. However, compatibility issues and the intrinsic complexity of migrating from Nvidia's well-established systems to domestic alternatives present substantial hurdles. According to insights shared by a government-backed think tank in Beijing, while there are ongoing developments in domestic technologies, the recommended path for Chinese data centres remains the use of Nvidia chips. This recommendation arises from the high costs and technical challenges associated with switching to local alternatives.
As pointed out in a recent report by the China Academy of Information and Communications Technology (CAICT), the complexities of this transition are significant. The report noted, "If the conditions allow, [data centres] can choose [Nvidia’s] A100 and H100 high-performance computing units. If the need for computing power is limited, they can also choose H20 or alternative domestic solutions." The report goes on to highlight the “trend of computing power fragmentation,” which has resulted in average GPU utilisation rates dropping below 40%. It suggests that discrepancies in hardware across internet data centres (IDCs) have further complicated resource management for AI tasks.
Despite these issues, China's AI computing power is witnessing substantial growth. As of 2023, the country's overall computing capacity, which encompasses central processing units (CPUs) and GPUs, surged by 27 per cent year-on-year, reaching 230 exaflops. Specifically, GPU capacity increased even more dramatically, growing by 70 per cent in the same time frame. This expansion has been facilitated by over 250 internet data centres being either completed or under construction by mid-2023 as part of an initiative to bolster “new infrastructure.”
Local governments, state-owned telecommunications firms, and private investors are backing this extensive infrastructure development. However, there are growing concerns around overcapacity and potential under-utilisation of these new facilities. The rapid pace of construction could outstrip the demand for computing power, resulting in inefficiencies that may hinder the effectiveness of China's AI ambitions in the long term.
In summary, while China continues to strive for self-sufficiency in GPU technology, rapid legislative changes and the complexities inherent in transitioning to new systems underline the challenges faced by the nation in its quest for technological independence in the AI sector.
Source: Noah Wire Services