GPU With 2,560 TOPS Chinese Team Develops A Novel Optical Chip

Related Vendors

FAULHABER Drive Systems

dataTec AG

Wuxi InfiMotion Technology Co., Ltd.

Due to US boycotts, China cannot purchase advanced chips like Nvidia's RTX 4090. Now it has developed an optical chip that is twice as fast as the US competition.

The chip, named "Meteor 1," is the "first highly parallel integrated optical computing chip," reports the South China Morning Post (SCMP) in Hong Kong. It can perform more than 100 computing processes in parallel and, purely computationally, achieves a peak performance of 2,560 TOPS, which is comparable to Nvidia's most advanced GPUs, writes the newspaper.

The new chip was developed by Xie Peng, Han Xilin, and their team at the "Shanghai Institute of Optics and Fine Mechanics" (SIOM), which is part of the Chinese Academy of Sciences, in cooperation with Hu Guangwei from the "Nanyang Technological University" (NTU) in Singapore.

The development enables "a hundredfold—and even higher—increase in optical computing power through ultra-high parallelism without increasing chip size, thereby opening up a novel technological path for future optical computers," the researchers write in their own article in the Chinese scientific journal eLight, dated June 17, 2025.

Teamwork

The scientists had previously split into several teams, each tackling different hurdles of computing with optical chips. The optical core chip they built, which operates not with electricity but with light, features a high transmission bandwidth (more than 40 nm) and thus enables numerous parallel computing processes.

It uses an "integrated microcavity frequency comb with an output spectrum of more than 80 nanometers and can support more than 200 wavelengths simultaneously," it states. Other components were also optimized, such as a driver board with more than 256 channels. Together, all these improvements enhanced the "computational parallelism" of the optical chip.

According to media reports, the team was able to set a "world record" by performing more than 100 computing tasks simultaneously on their system.

Breaking the Technical Barriers

Team leader Xie Peng earned his PhD from the Massachusetts Institute of Technology (MIT) in the USA and subsequently conducted research both in Oxford and at NTU in Singapore. It was only last year that he founded the team for the development of optical chips at SIOM in Shanghai. This research success in China is interesting for two different reasons.

On the one hand, it suggests a potential "leapfrogging" of China in the AI race with the USA. The GPUs from Nvidia, which are not supplied to China by Washington, are primarily relevant for AI data centers due to their parallel computing abilities. On the other hand, independent of political disputes, the development of optical computers in the USA, Europe, and China has recently encountered similar technical hurdles.

Relying on Parallelism

In two key parameters, the size of the chip matrix and the optical clock frequency, significant progress has already been made, and further optimizations are becoming increasingly difficult to achieve. This leaves the third important approach and solution: the parallelism of photonic information processing—and this is precisely where Xie and his team at SIOM in Shanghai are now focusing.

If their approach can be further refined in detail, it could potentially enable greater scaling of computational processes than the continued miniaturization of conventional GPUs or new methods like 3D packaging. "We are firmly convinced that optical computing, with the scalability that our approach potentially addresses, can meet the ever-growing demands for computational power driven by AI," the SCMP quotes the scientist. A whole range of new applications would thus become conceivable. (sb)

Future-Proof Authentication with Universal RFID Readers

Universal RFID Readers: One Solution for All Credentials