New Chinese AI Chip Firm Debuts with 3D Stacking to Bypass US Curbs

A Chinese artificial intelligence chip startup, led by industry veteran Wei Shaojun, has emerged from stealth operations. The company joins other firms, including Huawei Technologies, in using 3D stacking to navigate US technology export controls.

Credits: UNSPLASH

Dongfang Suanxin launched its corporate website and social media account. Wei Shaojun, who also serves as vice president of the China Semiconductor Industry Association, heads the company. This public debut presents Dongfang Suanxin as a significant participant in China's artificial intelligence computing industry.

And this public communication, the first since the company's establishment in 2024, indicates a broader push within China's chip sector. This push aims to create a self reliant artificial intelligence chip ecosystem. The objective is to bypass existing US technology export controls.

The company states it uses two main technologies: "software defined chips" and "3D stacked near memory computing". These are designed to deliver extremely high performance computing chips. The firm also confirmed its technology depends completely on a domestic supply chain.

This focus on 3D stacking aligns with a general trend in the global chip industry. After Moore's Law, major semiconductor companies increasingly adopt 3D chip architecture. This occurs as traditional scaling methods approach their physical constraints.

But the concept gained more traction in China following Huawei's recent proposal. The Big Tech firm introduced the "Tau Scaling Law". This new engineering principle seeks to achieve similar increases in transistor density through 3D architectural changes.

Dongfang Suanxin has launched products, including the DF1000 series. This series is a high performance artificial intelligence accelerator based on "3D near memory computing architecture". The DF1000 series offers standard interfaces for easy adaptation to server platforms from major domestic OEMs. It supports both training and inference for large artificial intelligence models, according to the company.

The startup maintains its headquarters in Shanghai's Zhangjiang high technology hub. It employs more than 500 people. Research and development departments operate in seven Chinese cities, including Beijing and Shenzhen, information from its website shows.

And the company has drawn significant investment. Initial backers include state backed entities such as the National Artificial Intelligence Industry Investment Fund. Jack Ma backed Yunfeng Capital, along with venture divisions of Xiaomi and JD.com, are also among these investors. This information comes from a Chinafund report.

Mr Ma founded Alibaba Group Holding, the parent company of the South China Morning Post.

Mr Wei, a professor at Tsinghua University's School of Integrated Circuits, has frequently expressed views on China's semiconductor strategies.

So, at an industry conference in Shenzhen recently, Mr Wei advocated for strong policies to develop skilled professionals. He stated that high end talent remains in short supply. He also pointed out that China's domestic chip industry shows considerable fragmentation. Eighty seven percent of chip design firms employ fewer than 100 staff.

Earlier this year, Mr Wei spoke to the state media outlet Global Times. He called on the domestic sector to stay "highly vigilant" and to maintain its commitment to "advancing domestic capabilities" in advanced technologies.

• Dongfang Suanxin, a Chinese artificial intelligence chip startup, has publicly launched its operations.

• The firm, led by Wei Shaojun, uses "software defined chips" and "3D stacked near memory computing" to create high performance chips, relying on a domestic supply chain.

• This strategy, shared with companies like Huawei, aims to bypass US technology export controls by focusing on advanced chip architecture.

• The company, based in Shanghai, employs over 500 staff and has secured investment from state backed entities, Yunfeng Capital, Xiaomi, and JD.com.

Source: SCMP