Huawei Atlas · Supernode / Cluster
Atlas 900 A3 SuperPoD
The Atlas 900 A3 SuperPoD serves large-scale AI data centers for internet, telecom and finance; through the Lingqu high-speed interconnect, up to 384 NPUs operate as one computer for LLM training and inference.
Key Highlights
- Up to 384 NPUs interconnected with 784GB/s bidirectional D2D
- 48TB unified-address on-chip memory
- non-blocking 1:1 bus bandwidth
- 200ns single-hop latency
- 307.2/288.7 PFLOPS FP16
- optical-electronic lossless fabric
- scalable logical supernodes of 16~384 cards
- liquid cooling for lower PUE.
Technical Specifications
| Form Factor | 12 compute cabinets (47U) + 4 bus cabinets (47U) |
|---|---|
| NPU | up to 384x Ascend 910 |
| CPU | up to 192x Kunpeng 920 |
| On-chip Memory | up to 384x128GB, up to 3.2TB/s |
| AI Compute | up to 307.2/288.7 PFLOPS FP16 |
| D2D Bandwidth | 784GB/s bidirectional |
| System Memory | up to 1536x DDR5 up to 5200MT/s |
| Interconnect | 灵衢(Lingqu) high-speed bus, 1:1 non-blocking |
| Cooling | liquid (compute) + air (bus) |
| FP8 or INT8 compute | 官网未公开 |
Typical Use Cases
Large-scale AI data centers for internet, telecom and financeLLM training and inference.
翻译稿；官网仅公开FP16算力，FP8/FP4/INT8整机算力官网未公开；官方亦称昇腾384超节点；img1: main; 资料链接为官网《Atlas 900 A3 SuperPoD超节点技术白皮书》落地页。
Related Products
Huawei Atlas
Atlas 200I DK A2 Developer Kit
Easy-to-use AI development kit for fast innovation View details →
Huawei Atlas
Atlas 200I A2 Acceleration Module
Edge AI accelerator module embedded in smart devices, robots and drones View details →
Huawei Atlas
Atlas 300I A2 Inference Card
High-performance AI inference card integrated into servers View details →
Huawei Atlas