




By utilizing the power of Huawei Ascend processors, our cards, servers, and clusters allow us to build diversified computing infrastructure solutions for different kinds of data centers. These solutions cover E2E process of deep learning inference and training.
• 256 - 1024 PFLOPS @FP16
• Interconnection of thousands of Ascend 910 processors
The cluster integrates HCCS, PCIe, and 100G RoCE high-speed interfaces, while also vertically integrating communication libraries, topologies, and low-latency networks. The linearity is greater than 80%.
• Full liquid cooling system, 46 kW of power consumption for each cabinet
• Liquid cooling ratio: > 95%; PUE: <>
• Equipment room footprint saved: 79%
• A single cabinet provides up to 20.48 PFLOPS FP16 computing power.
• The cluster can be scaled to 1 EFLOPS @FP16
• Industry's first full liquid cooling cluster with a PUE less than 1.15
• Saves 79% of the equipment room footprint, achieving higher cost-effectiveness
• Industry's fastest ResNet-50 training performance on ImageNet