As AI models get bigger, chipmakers are focusing on how to connect thousands of chips efficiently. This is more important than making individual processors faster.
These connected server systems are now a key area for AI companies. They are racing to build the computing power needed for huge AI models and more AI agents.
Breaking GPU Cluster Limits
Ding Yunfan, an AI architecture vice president at Biren, said that making a single graphics processing unit (GPU) faster is no longer enough. The biggest challenge is turning many GPUs into one smooth computing platform.
We're a new kind of news feed.
Regular news is designed to drain you. We're a non-profit built to restore you. Every story we publish is scored for impact, progress, and hope.
Start Your News DetoxBiren believes that future large-scale AI computing needs optical technology instead of traditional electrical connections. Ding explained that copper-based connections are reaching their limits. This caps current server systems at about 128 GPUs.
To fix this, Biren developed a new supernode architecture. It uses near-packaged optics (NPO), which places optical fibers closer to the chips. This increases bandwidth and improves data transfer. Biren says this method could allow clusters with up to 1,024 AI accelerator cards. This would greatly expand computing systems for demanding AI tasks.
Ding noted that optical connections are becoming vital to overcome the limits of current server designs. Shanghai-based Biren is one of many Chinese companies working on these infrastructure challenges. Other companies like MetaX, Enflame, and Alibaba are also developing supernode architectures. They aim to compete with Nvidia’s NVL72 and NVL144 AI platforms.
MetaX, a rival based in Shanghai, introduced its Xijing S600 AI supernode. It connects 64 GPUs in one rack. The company said its design reduces signal loss by removing external cables. This improves efficiency for large AI workloads.
Building Larger AI Supernodes with Optical Links
Biren is also exploring new ways to expand its AI computing systems. The company is testing a prototype NPO interconnect system. It is also working on an orthogonal hardware architecture with ZTE, which uses traditional electrical connections.
The chipmaker stated that the optical approach could eventually support supernodes with over 512 accelerator cards. This would help meet the growing demand for larger and more efficient AI infrastructure.
Despite the interest in optical networking for AI, Ding said the semiconductor industry is still in early stages. Widespread use will require more advancements before these systems can be deployed broadly.
Ding also stressed that the technology needs real-world testing. He expects NPO optical systems to be widely available around 2028. This gives the industry time to refine the technology and solve engineering issues.
