China unveils technology to interconnect up to 1,000 AI chips

China’s Biren Technology has unveiled an optical interconnect architecture capable of linking up to 1,024 AI accelerators into a single computing cluster, El.kz reports citing the South China Morning Post.

The solution is designed for deployment in data centers and for training large-scale AI models.

According to the publication, the company is leveraging optical data transmission to address the limitations of conventional electrical interconnects used in current server platforms.

The technology is based on Near-Packaged Optics (NPO), an architecture in which optical components are positioned in close proximity to computing chips. This approach is intended to increase data transfer bandwidth while reducing communication latency.

Such capabilities are particularly critical for AI systems, as large-scale AI models are trained and executed not on a single processor but across clusters comprising multiple graphics processing units (GPUs) and AI accelerators. The faster these processors exchange data, the more efficiently the overall computing infrastructure operates.

Biren Technology states that its new architecture is capable of supporting computing clusters of up to 1,024 AI accelerators. By comparison, the company estimates that existing server architectures based on electrical interconnects are typically limited to clusters of approximately 128 GPUs.

Optical interconnect technologies are increasingly recognized as a key direction in the development of next-generation AI infrastructure. Broadcom has previously noted that integrating optical components closer to processing chips can reduce power consumption and enable a greater share of system resources to be allocated to computational workloads.

Headquartered in Shanghai, Biren Technology develops graphics processing units (GPUs), AI accelerators, and server platforms. Amid restrictions on access to foreign semiconductor technologies, such developments form part of China’s broader strategy to strengthen its domestic AI infrastructure and technological self-reliance.