A concept picture of 5G chip Photo: VCG

A Chinese research team has developed a silicon photonic integrated high-order mode multiplexer chip, enabling ultra-high-capacity on-chip optical data transmission, Global Times learned from Fudan University on Thursday.

The breakthrough offers a new solution for optical interconnection in data centers and high-performance computing servers, while laying a solid foundation for artificial intelligence (AI), large-scale parallel computing, and model training. The findings have been published in Nature Communications.

As large language AI models scale up, communication bandwidth between intelligent computing chips and nodes is becoming a growing challenge. A research team from Fudan University's School of Information Science and Technology integrated multi-dimensional multiplexing technology into on-chip optical interconnects through precise design and optimization. 

This innovation significantly enhances data transmission throughput while achieving exceptional efficiency in power consumption and transmission latency, according to Science and Technology Daily's report.

With strong scalability and compatibility, the technology is applicable for various high-performance computing needs. Building on this, the team developed a silicon photonic integrated high-order mode multiplexer chip, enabling ultra-high-capacity on-chip optical data transmission. Tests show the chip supports 38 Tbps, capable of transferring 4.75 trillion model parameters per second, according to the report.

This advancement significantly enhances communication efficiency and reliability in large-scale model training and computing clusters, providing strong support for AI, model training, and GPU-accelerated computing, the newspaper report said.

"While most large-scale chips today are electronic, research into photonic chips has gained momentum, facilitating the shift from electronic to optical transmission," Ma Jihua, a veteran telecom industry observer, told the Global Times on Thursday. 

However, integrating optical and electronic components remains a challenge, as communication typically relies on CMOS technology. The new technology enables photonics-based interconnection, significantly reducing transmission latency and improving efficiency, according to Ma.

He highlighted that its high-order multiplexing greatly enhances transmission efficiency over traditional single-mode systems, aligning with the rising demand for high-speed, high-volume data transfer, especially in AI applications. The breakthrough boosts bandwidth and speeds, providing strong support for next-generation networks. 

Ma added that the new technology could greatly enhance domestic AI training, improving power use efficiency and commercial viability. "Major breakthroughs in application may come within the next three to five years," he said.

Global Times