Inside Huawei’s Atlas 850E & 950 Supernodes: Key Technical Innovations

Huawei’s Atlas 850E and 950 supernodes represent a strategic shift from single‑card performance to system‑level AI compute clusters, using self‑developed high‑speed interconnects and unified system architecture to treat thousands of compute units as a single logical supercomputer.

Atlas 850E Wind‑Cooled Supernode – High‑Performance Inference

The 850E is a 14U AI server that integrates eight Ascend 950DT NPUs and two Kunpeng CPUs, delivering up to 14.27 PFLOPS (mxFP4) of compute power. It features 768 GB of HBM memory with a bandwidth of 4.0 TB/s . Key innovations include:

VCE phase‑change cooling : enables stable operation in standard wind‑cooled cabinets at a maximum power draw of 15.6 kW without infrastructure upgrades.

Lingqu 2.0 high‑speed interconnect : provides a per‑node NPU interconnect bandwidth of 13.4 TB/s and can scale to a 96‑card inference cluster in a wind‑cooled environment.

The system architecture combines custom Kunpeng and Ascend 950DT processors, with a physical layout of one 14U chassis, one CPU drawer, and two NPU drawers. Logical connectivity links two Kunpeng 950 CPUs (each supporting 12 DDR DIMMs) and eight NPU modules via UB ports, forming a full‑mesh topology through PCIe and PCIe riser cards.

Atlas 950 SuperPoD – Flagship Training and High‑Concurrency Inference Platform

The 950 SuperPoD targets trillion‑parameter models, offering 1 EFLOPS FP8 compute and a cluster of 1 024 NPU cards (expandable to 8 192). Its three core technical advantages are:

Unified memory addressing : a global 256 TB address space eliminates memory walls for large models and long context lengths.

Ultra‑high bandwidth and ultra‑low latency : Lingqu 2.0 delivers 1.72 PB/s NPU interconnect bandwidth and 3 µs RTT , a 15× bandwidth increase and 10× latency reduction over traditional solutions.

Full liquid‑cooling architecture : system PUE as low as 1.15 solves power and thermal challenges for thousands of cards.

The logical architecture groups compute cabinets as minimal 64‑card nodes or larger clusters, with a maximum configuration of 128 cabinets (8 192 cards) occupying roughly 1 000 m². It integrates high‑density TPSU power supplies, advanced heat‑sink designs, and three‑dimensional orthogonal connectors to ensure reliable operation.

Both supernodes illustrate Huawei’s approach to building AI‑centric data center solutions that combine high compute density, scalable interconnects, and innovative cooling techniques to meet the demands of next‑generation AI workloads.