Huawei Atlas 960E SuperPoD: First NPO Super Node for 10-Trillion-Parameter Models

At the 2026 Huawei Connect conference on September 17, Huawei officially launched the Ascend 960 Super Node , marketed as the Atlas 960E SuperPoD . This system is the industry's first super node to adopt Near Package Optics (NPO) technology , built on the "Lingqu UnifiedBus + Hi-ONE" architecture to form a scalable super-node interconnect system designed to accelerate training and inference of models with ten-trillion parameters .

1. Atlas 960E Super Node Core Specifications and Performance

Compute and Scale: A single super node scales to 4,096 cards , delivering 8 EFLOPS FP8 and 16 EFLOPS FP4 compute throughput.

Architecture Design: Based on "Lingqu + Hi-ONE" , employing an orthogonal architecture with full liquid cooling and supporting unified memory addressing .

Optical Interconnect Breakthrough: Uses 5,500 Hi-ONE optical engines to replace the 48,000 800G optical modules otherwise required. This reduces power consumption by over 550 kW and doubles mean time between failures, achieving 99.8% availability .

Cluster Capability: Multiple super nodes interconnect via Lingqu network or RoCE to build clusters up to 512,000 cards , with a roadmap to 1 million cards in Ascend super-node clusters.

2. Ascend 950 SuperPoD Progress

Alongside the 960 launch, Huawei disclosed that the previous-generation Atlas 950 SuperPoD has entered large-scale commercial deployment . Its physical form comprises compute cabinets and Lingqu interconnect cabinets . A single cabinet supports 64 super nodes; the system scales to 1,024 super nodes. Over 1,000 Atlas 950 SuperPoD units have been deployed commercially , and the Ascend super node is noted as the only domestic super node that has trained SOTA (state-of-the-art) models .

3. Ascend Processor Release Schedule and Microarchitecture Details

Ascend 960DT — Ready Q1 2027 (three quarters ahead of plan)

Microarchitecture: SIMD/SIMT

Compute: 2 PFLOPS (FP8) / 4 PFLOPS (FP4)

Data formats: FP32, HF32, FP16, HiF4, etc.

Interconnect bandwidth: 2.2 TB/s

Memory: Capacity doubled to 288 GB , bandwidth 9.6 TB/s

Ascend 960PR — Ready Q3 2027 (one quarter ahead of plan)

Key change: FP4 compute raised to 8 PFLOPS

Trade-off: Memory capacity reduced to 192 GB , bandwidth to 2.4 TB/s

Ascend 970 — Planned 2028

Microarchitecture: SIMD/SIMT

Compute: 3.6 PFLOPS (FP8) / 14 PFLOPS (FP4)

Data formats: FP32, HF32, FP16, BF16, FP8, MXFP8, HiF8, HiF4

Memory: 288 GB capacity, interconnect bandwidth 4.4 TB/s

Ascend 980 — Planned 2029

Microarchitecture: SIMD/SIMT

Compute: 7.2 PFLOPS (FP8) / 28 PFLOPS (FP4)

Data formats: FP32, HF32, FP16, BF16, FP8, HiF4

Memory: 384 GB capacity, interconnect bandwidth 8 TB/s