Chinese chip start-up Biren bets on light-based ‘supernodes’ to match Nvidia
Chinese chip firm Biren Technology is developing "supernode" solutions using optical data transmission to connect thousands of AI chips, aiming to overcome hardware limitations and compete with Nvidia.
Intelligence analysis by Gemini 2.5 Flash Lite
Biren Technology is tackling the growing demand for AI computing power by introducing "supernode" systems that leverage near-packaged optics (NPO) to enable massive clusters of over 1,000 AI chips. This approach aims to bypass the physical limitations of traditional copper wiring, which restricts current server architectures, and provide a more scalable and seamless solution for advan…
Imagine building with LEGOs. Right now, the wires connecting the LEGO bricks are like thin strings that can only carry so much information at once. Biren is creating super-fast "light pipes" that connect many more LEGO bricks together, letting them share information much faster to build bigger and cooler things, like super-smart computer brains.
Analysis
The Bottleneck of Bandwidth
The relentless advancement of artificial intelligence, particularly with models scaling to trillions of parameters and the rise of AI agents, has created an insatiable appetite for computational power. However, the very hardware designed to accelerate these processes is encountering fundamental physical limitations. Traditional server architectures, reliant on copper wiring for electrical data transmission, are hitting a ceiling. Ding Yunfan, vice-president of AI framework architecture at Biren, highlighted this issue, noting that current systems are effectively capped at around 128 GPUs per server due to bandwidth constraints. This limitation hinders the seamless integration of massive GPU clusters, preventing them from operating as a unified, powerful system capable of handling the most demanding AI workloads.
Biren's Optical Solution: Supernodes and NPO
To circumvent these physical barriers, Biren Technology is pioneering a novel approach centered on "supernode" solutions. These systems are engineered to link thousands of AI chips into a single, cohesive cluster. The key innovation lies in their adoption of near-packaged optics (NPO). NPO integrates optical fibers much closer to the chips themselves, enabling significantly higher data transmission speeds and bypassing the limitations of electrical signaling. Biren's distributed, decoupled supernode architecture, powered by NPO, has the potential to scale up to an impressive 1,024 processor cards within a single cluster. This represents a substantial leap beyond current capabilities and positions Biren as a significant contender in the race to build the next generation of AI infrastructure.
The Shifting Landscape of AI Infrastructure
Biren's announcement at the World Artificial Intelligence Conference (WAIC) in Shanghai underscores a broader trend: the intense competition in the AI infrastructure space. Companies are no longer solely focused on individual chip performance but are increasingly emphasizing the architecture and interconnectivity of entire server clusters. The race is on to develop solutions that can efficiently scale computing power to meet the demands of ever-larger AI models and more sophisticated AI applications. By betting on optical interconnects, Biren is not only addressing a critical bottleneck but also signaling a potential paradigm shift in how AI supercomputing systems are designed and deployed, directly challenging the dominance of established players like Nvidia.
Key points
- Chinese chip start-up Biren Technology is developing "supernode" systems for AI infrastructure.
- These systems utilize near-packaged optics (NPO) to enable high-speed data transmission between thousands of AI chips.
- The goal is to overcome the limitations of traditional copper wiring in large GPU clusters.
- Biren's architecture aims to scale up to 1,024 processor cards per cluster, surpassing current hardware limits.
- This innovation positions Biren as a competitor in the rapidly evolving AI infrastructure market.
If Biren's light-based supernode technology proves successful, it could significantly accelerate the development and deployment of more powerful AI models and applications. This advancement might lead to breakthroughs in fields requiring massive computational power, such as scientific research, drug discovery, and complex simulations, while also fostering greater competition in the AI hardware market.
The transition to optical interconnects is complex and costly, and Biren faces significant challenges in scaling production and ensuring reliability against established competitors like Nvidia. If the technology fails to meet performance expectations or incurs prohibitive costs, Biren could struggle to gain market traction, leaving the AI infrastructure landscape dominated by existing solutions.
