Skip to content
Huawei Atlas 850E: Air-Cooled AI Supernode Brings 96-NPU Scale to Standard Data Centers

Huawei Atlas 850E: Air-Cooled AI Supernode Brings 96-NPU Scale to Standard Data Centers

Huawei's Atlas 850E is a 14U air-cooled AI supernode integrating eight Ascend 950DT NPUs and two Kunpeng 950 CPUs, delivering 14.27 PFLOPS at mxFP4 with 768GB HBM and 13.4 TB/s internal interconnect, using VCE phase-change cooling to enable 15.6kW deployment in standard 19-inch racks without liquid cooling infrastructure.

Architects' Tech Alliance
Architects' Tech Alliance
Architects' Tech Alliance

Key Technical Innovations

The Atlas 850E introduces four core innovations that enable supernode-class compute in conventional air-cooled facilities:

VCE Phase-Change Cooling : Custom vapor-chamber evaporative cooling allows the 15.6 kW chassis to operate stably in standard air-cooled cabinets without any IDC infrastructure retrofit.

Air-Cooled High-Density Design : Eight Ascend 950DT NPUs and two Kunpeng 950 CPUs are packed into a 14U enclosure cooled by 30 hot-swap fan modules (N+1 redundancy), supporting direct deployment in standard 19-inch racks.

High-Speed Interconnect (Lianqui 2.0) : The UB protocol implements a Full Mesh topology among the eight NPUs inside the chassis, delivering 13.4 TB/s total NPU-to-NPU bandwidth (1.68 TB/s per NPU bidirectional). The design scales to a 96-NPU inference cluster in an air-cooled environment.

Low-Precision Formats with High-Bandwidth HBM : Each Ascend 950DT NPU natively supports mxFP4, mxFP8, and HiF8 formats, reducing precision loss while boosting inference throughput. Paired with 96 GB HBM per NPU (768 GB total, 4.0 TB/s aggregate bandwidth), the system achieves stable sub-10 ms latency for large-model and long-context workloads.

Detailed Specifications

Form Factor : 14U AI air-cooled server

Processors : 8× Ascend 950DT NPU + 2× Kunpeng 950 CPU (up to 96 cores, 2.3 GHz)

AI Compute : 14.27 PFLOPS @ mxFP4 / 7.35 PFLOPS @ FP8 / 3.88 PFLOPS @ FP16 / 1.944 PFLOPS @ HF32 / 0.216 PFLOPS @ FP32

On-Chip Memory (HBM) : 8× 96 GB (≥768 GB total), max bandwidth 4.0 TB/s

System Memory : 24 DDR5 slots, up to 6400 MT/s, max 2304 GB

NPU Scale : Standard 96-NPU configuration; flexible expansion from 8 NPUs to thousand-NPU level (768 or 1024 NPUs)

NPU Interconnect : Total 13.4 TB/s (8 × 1.68 TB/s bidirectional, Lianqui 2.0)

External Network : 400 Gbps per NPU (UBoE or RoCE)

Expansion Slots : Up to 5× PCIe 5.0 slots

Power Supply : 6× 3.0 kW hot-swap PSUs, 5+1 redundancy, 220 V AC / 336 V HVDC

Cooling : 30 hot-swap fan modules, N+1 redundancy, air-cooled

Max Power : 15.6 kW

Dimensions (H×W×D) : 618.8 mm × 447 mm × 920 mm, net weight 251 kg

Cluster Expansion : Supports Lianqui-based 96 or 768 NPU air-cooled AI supernodes

System Architecture

Physical Structure

The 14U chassis adopts an air-cooled horizontal plug-in architecture. All drawers support power-off insertion/removal, reducing maintenance complexity. The frame provides structural support and distributes 54 V DC to NPU and CPU drawers via a power transfer board and busbars.

Internal layout:

1× 14U chassis

1× CPU drawer (3U)

2× NPU drawers (4.5U each)

Maximum input power 15.6 kW, fed by six 3.0 kW PSU modules with 5+1 redundancy.

Component Classification

Hot-Pluggable Modules : NPU drawers, CPU drawer, 30 fan modules, 6 PSU modules. NPU and CPU drawers support whole-drawer cold-plug; fans and PSUs support hot-plug.

Frame-Integrated Components : Power transfer board, power management board, fan backplane, busbars, cable backplane, PSU backplane, horizontal copper bars — all fixed to the frame as a single assembly.

The CableTray is the key internal interconnect module, routing high-speed cables between NPUs, between NPUs and CPUs, and low-speed NPU management signals.

Logical Structure

Two Kunpeng 950 CPUs, each with 12 DDR DIMM slots.

Eight NPU modules; each NPU module exposes 8 internal UB ports plus 1 parameter-side port for AI processor interconnect.

CPUs connect to the eight NPUs via 2 UB ports each, forming a direct CPU–NPU link.

The eight NPUs form a Full Mesh interconnect.

Two CPUs interconnect via PCIe bus to PCIe devices; different PCIe riser cards support varying PCIe slot specifications.

Compute & Memory Detail

Each NPU delivers up to 1784 TFLOPS @ mxFP4. Eight NPUs aggregate to the peak figures listed above. Each NPU provides 96 GB HBM at 4.0 TB/s, totaling 768 GB on-chip memory. NPU interconnect: 15 UB ports per module, 1.68 TB/s per port (bidirectional), yielding 13.4 TB/s chassis-wide.

Two Kunpeng 950 CPUs, 96 cores each at 2.3 GHz. System memory: 24 DIMM slots, up to 6400 MT/s, 96 GB per DIMM max. CPUs handle data preprocessing, task scheduling, and mixed compute loads.

Comparison: Atlas 850E vs. Atlas 950 / CloudMatrix 384

Fundamental Difference : Air cooling vs. liquid cooling. Atlas 850E, via proprietary VCE phase-change cooling, is the first supernode deployable in general-purpose air-cooled data centers, bringing supernode technology to a broader set of enterprises.

Scale & Performance Gap : Atlas 950 SuperPoD targets a single scale-up domain of 8192 NPUs, serving as a future "ten-thousand-NPU" foundation. Atlas 850E's 96-NPU scale is comparatively lightweight.

Product Evolution : Huawei's supernode lineage progressed from Atlas 900 A3 validation, to Atlas 950/960 (thousand/ten-thousand NPU flagship), and now to Atlas 850E (air-cooled, 96 NPU) for commercialized mainstream adoption.

Original Source

Signed-in readers can open the original source through BestHub's protected redirect.

Sign in to view source
Republication Notice

This article has been distilled and summarized from source material, then republished for learning and reference. If you believe it infringes your rights, please contactadmin@besthub.devand we will review it promptly.

AI supernodeHBM memoryKunpeng 950Atlas 850Eair-cooled data centerAscend 950DTLianqui 2.0VCE cooling
Reader feedback

How this landed with the community

Sign in to like

Rate this article

Was this worth your time?

Sign in to rate
Discussion

0 Comments

Thoughtful readers leave field notes, pushback, and hard-won operational detail here.