[Tech Talk] China's AI Counterattack?..."Developing NVIDIA-level GPUs"

"A100 and AI Training Time Almost Identical"
Huawei Ascend 910, Nvidia Alternative
China, Restricted from Export, Successfully Develops Domestic GPU

"Huawei has developed a GPU comparable to NVIDIA's A100."

This is recent news that has stirred up the Chinese tech community. Huawei, a leading Chinese IT company, reportedly developed an artificial intelligence (AI) accelerator on par with NVIDIA's flagship data center GPU, the A100.

The biggest obstacle to the development of China's AI industry is the lack of domestically produced GPUs. Ultimately, GPUs must be imported from overseas companies, but active technology export restrictions imposed by the U.S. government have caused disruptions. This is why the global spotlight is on the actual performance of Huawei's GPU.

A New Semiconductor Powerhouse in the Continent? ... Huawei Successfully Develops A100-Level GPU

Ascend 910 GPU, a Chinese AI semiconductor designed by Huawei. [Image source=Huawei]

This claim came from the Chinese tech company iFlytek on the 27th (local time). They stated that training a massive AI model on a data center GPU provided by Huawei took a similar amount of time as training on the A100. If the company's claim is true, Huawei's GPU likely possesses performance comparable to the A100.

However, this alone does not mean that China's AI industry has established a self-sufficient ecosystem. This is because the A100 GPU is already a previous-generation product. NVIDIA has already released the next-generation H100 GPU architecture this year, and plans to launch an AI-specialized computer system called GC200 next year.

The H100 offers nine times higher processing performance compared to the A100, and the GC200 is expected to achieve even higher performance with the help of next-generation Grace superchips and HBM3e memory.

Moreover, several AI accelerator startups chasing NVIDIA (such as Intel Habana, Graphcore, Cerebras, SambaNova, etc.) have already developed GPUs with A100-level performance and plan to release next-generation chips that will catch up to the H100 between the second half of this year and the first half of next year. In terms of hardware performance competition, China is still one generation behind the West.

Chinese Semiconductor Companies Begin Building 'Ecosystems' Like NVIDIA

Huawei is leading China's self-reliance in advanced computer chips.
[Image source=Reuters Yonhap News]

However, even if the performance of Chinese GPUs is currently lower, there are still reasons to be cautious. The Huawei GPU in question, the 'Ascend 910,' currently holds a significant market share in the Chinese GPU market. Additionally, many Chinese developers are already using Ascend's SDK (Software Development Kit).

The reason NVIDIA has been able to almost monopolize the AI semiconductor market is due to the 'CUDA' toolkit. First released in 2007, it has been the primary factor in attracting countless developers into the NVIDIA ecosystem.

An SDK is a collection of development tools used by developers to create software. Developers who have become accustomed to the CUDA toolkit over the past 16 years prefer to stay with CUDA rather than learning new SDKs for other GPUs or AI accelerators. This is why NVIDIA has been able to thrive without losing customers to competitors.

However, the situation is different in the Chinese market. Last September, the U.S. Department of Commerce restricted the export of NVIDIA's A100 and H100 (then unreleased) GPUs to China. Only versions of existing GPUs with reduced performance were freely allowed for export to China.

As a result, Chinese developers were forced to find alternatives to NVIDIA GPUs. Consequently, various GPU companies, including Huawei's Ascend 910, have emerged in China, and more importantly, developers are leaving CUDA.

Of course, in terms of performance, Chinese GPUs may still lag behind NVIDIA's, but the fact that Chinese companies are expanding their own development ecosystems is an undeniable strength.

NVIDIA Founder: "Chinese Companies Now Make Chips Locally"

Jensen Huang, CEO of NVIDIA
Photo by AFP Yonhap News

Whether the 'advanced technology barrier' directly imposed by the U.S. on China will break China's ambition for technological advancement or instead become a starting point for China to become a self-sufficient tech powerhouse is something no one can predict.

However, Jensen Huang, founder and CEO of NVIDIA, warned in an interview with the Financial Times (FT) in May that U.S. export restrictions could cause "tremendous damage."

At the time, Huang explained, "China already accounts for about one-third of the U.S. technology industry's market," and added, "If we distance ourselves from the Chinese market, U.S. semiconductor manufacturing facilities will become overproduced."

He also expressed concern that the export restrictions on advanced computer chips could result in a windfall for China's semiconductor industry. Huang emphasized, "Due to the ban, our company is tied behind the back when trading advanced chips with China," and "Chinese companies cut off from U.S. firms are making chips locally."