Biren Tech unveils supernode solution supporting 1024 GPUs

At the 2026 World AI Conference, Biren Tech launched a new supernode solution with NPO optical interconnect and distributed decoupled architecture, supporting up to 1024 GPUs per node. It's built on the company's BR2xx series GPUs and BLink2.0 protocol, allowing all 1024 GPUs to share the same memory space.

The BR2xx GPUs use Biren's innovative Chiplet architecture, support low-precision FP8/FP4 computing, and have larger memory bandwidth with native supernode interconnect. Biren created three product tiers: a 16-card standard server supernode, a 128-card high-density rack supernode, and a 1024-card distributed decoupled supernode.

The 16 and 128 card versions target small to medium clients with billion to trillion parameter models, while the 1024-card NPO optical solution is for large clients with multi-trillion parameter scenarios, offering greater scalability for AI computing.