Alibaba Cloud's Panjiu Predictable Network: High‑Performance Architecture and Core Technologies

At the inaugural China Computing Power Conference, Alibaba Cloud highlighted its Panjiu Predictable Network, a self‑designed high‑performance data‑center network built around an application‑centric approach and the "Alibaba Cloud full‑stack self‑development + end‑network integration" technology.

The core of the solution is the High Performance Network (HPN) architecture, a dual‑plane, two‑level Clos design without convergence that can support clusters of more than 10,000 A100 GPUs, delivering minimal static forwarding latency and low congestion probability.

Dual‑plane design ensures that a failure on a single device or plane does not affect the whole cluster, while a double‑uplink service access further enhances stability and reliability.

All network devices, optical interconnects, and the high‑performance protocol stack are self‑developed. The AliNOS software unifies device‑level and network‑level monitoring, enabling rapid feature iteration and integrated control.

Alibaba Cloud also created its own high‑performance protocol, Solar‑RDMA, which replaces IB and RoCEv2. Solar‑RDMA, combined with the HPCC congestion‑control algorithm, eliminates the need for PFC, offering high bandwidth, low latency, and stable large‑scale deployment.

To accelerate Solar‑RDMA, a dedicated hardware offload was designed, resulting in the Fusion Intelligence Card (FIC) – a self‑developed high‑performance NIC that has been rolled out at scale.

On the service side, the Network Unified Service Architecture (NUSA) provides end‑to‑end automation for network provisioning, performance measurement, fault monitoring, resource management, and virtualization, delivering a trustworthy, cost‑effective high‑performance network experience for customers.

Through these innovations, Alibaba Cloud aims to usher in a new era of predictable data‑center networking, continuously evolving toward richer communication semantics, higher bandwidth, lower latency, and better usability.