Servers
AI servers, your compute powerhouse
Atlas 650E Server
Atlas 650E Server
The Atlas 650E server, powered by eight 950DT NPUs, delivers up to 12.4 PFLOPS @ mxFP4. It features Huawei’s self-developed on-chip memory with up to 4.0 TB/s R/W bandwidth while supporting 16-NPU full-mesh via UnifiedBus Link 2.0 protocol. Designed for standard air-cooled data centers, it offers efficient, reliable AI computing for industries such as finance, government, healthcare, etc.
Product Features
- Achieve industry-leading compute performance of 12.4 PFLOPS @ mxFP4
- Supports HiF8/mxFP8/mxFP4 data formats, significantly improving quantization acceleration performance
- Industry-leading 4.0 TB/s HBM read/write bandwidth
- 768 GB HBM capacity per server
- Supports up to 13.4 TB/s NPU interconnect bandwidth per server (dual-node scenario)
- Supports 16-NPU full-mesh interconnect between two nodes based on LingQu
Atlas 650E Server Specifications
 | |
Form | 14U rack server |
NPU | 8 x Ascend 950DT |
CPU | 2 x Kunpeng 950 |
*AI Computing Power | 12.48 PFLOPS mxFP4 6.43 PFLOPS mxFP8/FP8/HiF8 3.40 PFLOPS FP16/BF16 |
* Theoretical computing power may have a deviation of less than 1% in actual tests | Theoretical computing power may have a deviation of less than 1% in actual tests. |
On-chip Memory | 8 x 96 GB HBM, memory bandwidth up to 4.0 TB/s 768 GB HBM capacity per server |
System Memory | Supports 24 memory slots running at up to 6,400 MT/s |
Local Storage | 8 x 2.5 NVMe + 2 x 2.5 SATA (with RAID card configured) Or 8 x 2.5 NVMe |
NPU LingQu Interconnect | Single-node total bandwidth: 8 x 784 GB/s (bidirectional) Dual-node total bandwidth: 2 x 8 x 1.68 TB/s (bidirectional) |
Network | UBoE: Directly routed from NPUs, 400 Gbps per NPU (unidirectional) |
PCIe Ports | Up to 5 PCIe 5.0 expansion slots |
PSU | 6 x hot-swappable 3.0 kW PSU modules, supporting 5+1 redundancy Supporting hot swapping and 5+1 redundancy |
Power Supply | 220 V AC |
Heat Dissipation | Air cooling |
Fan Modules | 30 fan modules integrated per server, supporting hot swapping and N+1 redundancy |
Power Usage | Maximum input power consumption of 14.5 kW |
Operating Temperature | 5°C to 35°C (41°F to 95°F) |
Dimension | 618.8 mm (H) x 447 mm (W) x 920 mm (D) |
Application Scenarios
Intelligent Agent
AI Coding
Assisted Office
Content Review
Financial Risk Control
Intelligent Customer Service