Biren's 1024-GPU Super Node: NPO Optical Interconnect & Chiplet Architecture Explained

BR20X Series GPU: Next-Generation Training Chip

Biren's next-generation flagship training chip, the BR20X series GPU, introduces native support for FP8 and FP4 low-precision formats to boost large-model training and inference efficiency. Compared with the first generation, it features larger-capacity, higher-speed memory and higher interconnect bandwidth to alleviate data-movement bottlenecks.

Interconnect & Scale-Up

The BR20X series integrates super-node interconnect capability natively, relying on the proprietary BLink 2.0 super-node interconnect protocol to connect up to 1024 GPUs in a single scale-up domain. The accompanying system design, "LightSphere X" , adopts NPO (Near-Package Optics) optical interconnect and a distributed decoupled architecture .

Memory-semantic interconnect: Enables up to 1024 GPUs to share a single memory space, operating as one "super GPU."

In-network computing: Offloads mathematical operations from communication to the switch, reducing GPU compute load.

Intelligent congestion control & link self-healing: Multi-layer protection from physical to framework layer to prevent network congestion and ensure stability for large-scale training jobs.

Three-Tier Super Node Product Matrix

Based on the BR2xx chips and BLink 2.0, Biren defines three product tiers:

Standard Server Super Node — Max scale: 16 GPUs; Interconnect: Electrical; Target scenario: SMEs, hundred-billion-parameter models.

High-Density Cabinet Super Node — Max scale: 128 GPUs; Interconnect: Electrical; Target scenario: Scaled deployment of hundred-billion to trillion-parameter models.

Distributed Decoupled Architecture Super Node — Max scale: 1024 GPUs; Interconnect: NPO Optical; Target scenario: Large customers, ten-trillion-parameter models; NPO optical interconnect and large-scale expansion suited for hyperscale scenarios.

1024-Card Distributed Decoupled Architecture (NPO Optical)

The flagship 1024-card solution physically separates GPU nodes from switch nodes, allowing "Lego-like" expansion to 1024 cards. It is based on the next-generation BR20X series GPU , with formal launch expected in H2 2026 , while large-scale commercial deployment of NPO optical interconnect is projected for 2028 . The roadmap also mentions BR30X series planned through 2028 .