Huawei AI Server Series

Huawei's AI server series includes the Atlas and TaiShan lines, powered by Ascend NPUs, designed for large-scale AI training and inference with SuperPoD architecture.

Overview of Huawei AI Servers

Huawei's AI servers are built around Ascend NPUs (Neural Processing Units), which are purpose-designed for AI workloads rather than adapted from graphics processing. The main server series include:

• Atlas 950 SuperPoD: Integrates up to 8,192 Ascend 950DT NPUs per SuperPoD, delivering up to 16 EFLOPS FP16 and 8 EFLOPS FP8 performance. It uses Huawei's UnifiedBus interconnect to operate thousands of NPUs as a single logical machine, enabling large-scale AI training and high-concurrency inference .
• Atlas 960 SuperPoD: A next-generation cluster designed to scale even further, supporting up to 500,000–1 million Ascend NPUs in a SuperCluster configuration, targeting the largest AI compute workloads globally .
• TaiShan 950 SuperPoD: Focused on general-purpose AI computing with low-latency memory pooling and high bandwidth, suitable for both training and inference in carrier and enterprise environments .
• TaiShan 500 and TaiShan 200 series: Cover medium and low computing power requirements, providing flexible deployment options for diverse AI workloads .
• Atlas 850E: Supports air-cooled data centers, scaling from 8 to 1,024 NPUs, ideal for smaller-scale inference clusters .

Ascend NPU Roadmap

Huawei's AI servers are powered by the Ascend series of NPUs, with a roadmap spanning multiple generations:

• Ascend 950 series (2026): Includes 950PR for inference and recommendation workloads and 950DT for training. Features new low-precision formats (FP8, MXFP8, MXFP4, HiF8) and high-bandwidth memory (HBM) for efficient AI computation .
• Ascend 960 series (2027): Expected to double compute capacity over the 950 series, targeting larger AI models and higher throughput.
• Ascend 970 series (2028): Planned for further scaling, continuing Huawei's focus on high-performance AI training and inference .

Key Features

• SuperPoD Architecture: Combines multiple cabinets into a single logical machine with unified memory addressing and low-latency interconnect, enabling massive parallelism .
• High Scalability: Servers can scale from small clusters to SuperClusters with hundreds of thousands of NPUs.
• Optimized for AI Workloads: Da Vinci architecture in Ascend NPUs is optimized for tensor operations, matrix computation, and large-scale model training.
• Flexible Deployment: Options for air-cooled or high-density data centers, supporting both inference and training workloads. Huawei's AI server series represents a comprehensive ecosystem for AI infrastructure, from small-scale inference to the world's largest AI clusters, providing an alternative to NVIDIA's GPU-based solutions while leveraging domestic chip development and proprietary interconnect technology .