China’s Biren Targets 1,024-Card Scale-Up With NPO Optical Supernode

Biren Technology introduced an NPO optical interconnect and distributed, decoupled supernode architecture at last week’s WAIC 2026 in Shanghai, targeting scale-up connectivity for as many as 1,024 GPU accelerator cards. The Shanghai-based company positioned the design as the optical tier of a three-level product portfolio that also includes a 16-card standard server and a 128-card high-density rack, both based on electrical interconnects.

The proposed system combines Biren’s chiplet-based BR2xx GPU generation with its BLink 2.0 interconnect protocol. In public presentations, Biren said the GPUs support FP8 and FP4 computation, increased memory bandwidth, and integrated supernode connectivity. BLink 2.0 provides memory-semantic interconnects, in-network computation, congestion control, and multilayer link recovery, with Biren claiming that as many as 1,024 GPUs can operate within a unified memory space. The NPO design places optical engines close to the GPU modules and omits a separate high-power DSP stage, while fiber links physically separate the GPU nodes from the switching nodes.

Biren also described a “Token Factory” software architecture that uses five caching tiers and reportedly delivers a KV-cache hit rate above 95%. A separate heterogeneous inference project with China Telecom reportedly increased effective throughput by 20%, although Biren did not identify the participating accelerator platforms or establish that different GPU types can operate inside the same BLink scale-up domain. The company’s 2025 financial results separately confirm deployment of a 2,048-card GPU cluster using optical interconnect and distributed optical circuit switching at a national-level computing platform.

• Maximum proposed scale-up domain: 1,024 GPU accelerator cards

• Optical architecture: Near-Packaged Optics with physically separated GPU and switch nodes

• Electrical product tiers: 16-card standard server and 128-card rack

• GPU generation: BR2xx, based on a chiplet architecture with FP8 and FP4 support

• Interconnect: BLink 2.0 with memory semantics, in-network computing, congestion control, and link recovery

• Software: Five-tier KV-cache hierarchy with a claimed hit rate above 95%

• Previous optical deployment: 2,048-card cluster using optical interconnect and dOCS

• Commercial status: No customer, pricing, production volume, or general-availability date disclosed for the 1,024-card NPO system

“Optical interconnect has become the inevitable choice to break through this bottleneck,” said Ding Yunfan, Biren’s vice president of AI framework architecture.