Biren unveils 1,024-accelerator optical AI supernode

Biren Technology has unveiled an AI supernode architecture designed to combine up to 1,024 accelerators in a single system. The company presented it at the World Artificial Intelligence Conference (WAIC) in Shanghai, arguing that optical interconnects can overcome limits emerging in large AI clusters.

As accelerator servers approach a practical ceiling of roughly 128 GPUs, the bottleneck increasingly shifts from computation to communication: moving data quickly and efficiently between neighboring chips. Biren’s design uses optical links to replace some copper connections while bringing the compute units into a shared memory space.

Biren’s three supernode configurations

The company presented a family of BR2xx accelerators for these systems, along with Blink 2.0, a protocol intended to create a common address space across large numbers of cards. That approach could move some of the complexity of distributed computing from software into the infrastructure layer.

Recommended reading

Summarization framework cuts compute by up to 40%

Biren described three system tiers:

• 16 accelerators for standard servers
• 128 accelerators for high-density configurations
• A distributed optical supernode with 1,024 accelerators

The largest configuration relies on optics to reduce the limitations imposed by cable length, distance between nodes, signal loss, and the energy required to transfer data.

The architecture also uses near-packaged optics (NPO), positioned close to the computing components. NPO acts as an intermediate layer between the compute block and the network, shortening the electrical paths and reducing losses associated with long copper connections.

Optical interconnects move up the AI infrastructure stack

Biren is not alone in pursuing optical architectures. Chinese companies including Huawei, Alibaba, MetaX, and Enflame are developing similar approaches, while companies outside China are also studying optical systems as conventional designs become harder to scale.

When individual accelerators become faster, overall performance depends increasingly on how quickly hundreds of devices can exchange tensors and synchronize. Even small network delays can translate into lost training time, while larger clusters also raise demands for memory, cooling, and software support.

Optics could become a new standard if manufacturers can deliver the required performance, software compatibility, and production costs. Independent testing will determine whether Biren’s architecture progresses beyond a demonstration into data centers; at an affordable price, competition could shift from individual accelerators to the fastest and densest supernodes.