Huawei s AI Server

Huawei's AI server system, led by the CloudMatrix 384 and Ascend NPUs, delivers high-density, scalable AI compute with up to 300 petaflops of performance, leveraging proprietary interconnects and a full-stack AI ecosystem. Architecture and HardwareHuawei's AI servers are built around Ascend NPUs (Neural Processing Units), designed specifically for AI workloads rather than adapted from graphics pipelines. The current lineup includes the Ascend 910C and upcoming Ascend 950DT, with a roadmap extending to the 960 and 970 series.

These NP

Us use the Da Vinci architecture, optimized for matrix and tensor operations, enabling efficient neural network training and inference.

The Cloud

Matrix 384 system integrates 384 Ascend 910C chips in a “supernode” configuration, achieving 300 petaflops of BF16 compute performance, which is roughly double the compute of Nvidia's GB200 NVL72 system. The design emphasizes high-speed optical interconnects and low-latency communication between NPUs, allowing the cluster to operate as a single logical machine.

Scalability and Interconnect

Huawei employs the UnifiedBus (UB) 2. 0 protocol, a proprietary memory fabric that links thousands of NPUs in a SuperPoD or even a SuperCluster. This allows unified memory addressing and high-bandwidth, low-latency communication across all nodes, fundamentally improving scaling compared to traditional horizontal scale-out architectures.

The system can connect up to 8,192 Ascend 950DT NPUs in a single SuperPoD, with future SuperClusters projected to reach millions of NPUs.

Software and Ecosystem

Huawei provides a full-stack AI ecosystem including MindSpore, CANN, and openEuler, offering an alternative to Nvidia's CUDA. This stack supports real-time inferencing, intelligent scheduling, and context-aware services, and is optimized for joint deployment with Huawei's hardware. The open-sourcing of UnifiedBus 2.

0 encourages ecosystem growth and compatibility with third-party components.

Performance and Efficiency

While the CloudMatrix 384 delivers exceptional compute density, it is power-intensive, consuming 559 kW, nearly four times the energy of comparable Nvidia systems. The system is priced at around $8 million, reflecting its focus on providing a domestic AI compute resource for China rather than global market competitiveness. Huawei is also developing the Ascend 910D, which aims to improve power efficiency while maintaining high performance.

Enterprise and A

I ApplicationsHuawei's AI servers are designed for enterprise AI adoption, including large-scale training, inference, and cloud integration.

The Tai

Shan 950 SuperPoD complements the Atlas SuperPoD nodes for general-purpose workloads with AI inferencing support. The vertically integrated approach—from chip design to operating system and O&M software—enables optimized performance and reliability for large-scale AI deployments. SummaryHuawei's AI server system represents a highly integrated, scalable, and AI-optimized infrastructure.

Its strengths include massive compute density, proprietary interconnects, and a full-stack AI ecosystem, making it a viable alternative to Nvidia for domestic AI development. Trade-offs include high power consumption and cost, but the system is strategically positioned to support China's AI ambitions and enterprise-scale AI workloads.