Huawei Atlas 800I A3 8NVMe+2SFF
- Condition
- New
- Manufacturer
- Huawei
- CPU
- 4x Kunpeng 920 (64c/64t, 2.6Ghz, 180W)
- RAM
- 2000GB (DDR5 ECC REG)
- NIC card
- 8x 400GE QSFP-DD + 56x 400GE QSFP-DD
- GPU
- 8 x Ascend 910C
- Accelerator manufacturer
- Huawei
- Accelerator model
- Huawei Ascend 910C NPU
- Installed GPU/APU count
- 8 NPUs
- Maximum GPU/APU count
- 8 NPUs per server; up to 384 NPUs in a supported supernode
- Memory per accelerator
- 128GB
- Total accelerator memory
- 1024GB
- GPU memory type
- HBM
- Accelerator memory bandwidth
- 3.2TB/s per NPU
- Accelerator form factor
- Integrated supernode NPU modules
- Accelerator interconnect
- Huawei high-speed D2D supernode fabric, 784GB/s bidirectional bandwidth per NPU
- Accelerator-to-CPU interface
- PCIe 5.0
- Supported compute data types
- FP16 and INT8
- Accelerator virtualization
- Ascend device passthrough and containerized deployment through the CANN software stack
- MIG support
- Not applicable; NVIDIA MIG is not used on Huawei Ascend accelerators
- Software ecosystem
- Huawei CANN, MindSpore, MindIE, AscendCL and the PyTorch Adapter for Ascend
- Maximum system power
- 14.6kW
- Cooling
- Air cooling; each NPU drawer uses 5 hot-swap fan modules with 4+1 redundancy
- Recommended workloads
- Large-model inference, high-concurrency generative AI services, retrieval-augmented generation, enterprise knowledge systems, financial AI and supernode-scale distributed inference
- Dimensions
- 442 x 447 x 920 mm (H x W x D)
Need a technical consultation on choosing a server?
Submit a requestHuawei Atlas 800I A3 8NVMe+2SFF is a high-density 10U enterprise AI inference supernode storage configuration with eight 2.5-inch NVMe bays plus two 2.5-inch SATA SFF bays. It is designed for large enterprises, cloud operators and research organizations that need high-throughput large-model inference and supernode-scale expansion.
Key features:
-
Storage basket and service access: The configuration provides eight 2.5-inch NVMe bays plus two 2.5-inch SATA SFF bays. The 8+2 layout gives the high-performance NVMe tier most of the front storage positions while retaining two SATA SFF bays for capacity or system data.
-
Storage topology: This basket supports up to 8 x 2.5-inch NVMe drives and 2 x 2.5-inch SATA drives, matching the official maximum local-storage topology for Atlas 800I A3.
-
Installed accelerator subsystem: The installed subsystem contains 8 Ascend 910C accelerators for large-model inference and high-concurrency AI services.
-
Enterprise compute platform: The compute subsystem uses 4 Kunpeng 920 processors and 2,000GB of DDR5 ECC registered memory within the platform’s 32-DIMM architecture.
-
Management, networking and availability: Eight external 400GE QSFP-DD RoCE ports, the high-speed supernode fabric and redundant hot-swap fan modules support distributed inference, fast collective communication and service continuity.
Use cases:
-
Large-model inference: Runs generative AI, LLM and multimodal inference with the installed Ascend accelerator subsystem.
-
Enterprise content and knowledge services: Supports intelligent search, content generation, knowledge retrieval and customer-service automation.
-
Distributed inference clusters: Uses high-speed RoCE connectivity to scale inference services across multiple nodes.
-
Data staging and model serving: Uses the defined SATA and NVMe tiers for operating data, models, caches and service logs.
Huawei Atlas 800I A3 8NVMe+2SFF is a purpose-built storage configuration for organizations that need eight NVMe bays plus two SATA SFF bays, 8 installed Ascend 910C accelerators and high-speed supernode connectivity for large-model inference.
to Buy a Server
from Servermall
warranty
testing
We want to make sure that our equipment is fully reliable for you. That’s why all our equipment (even new) is subject to complete diagnostics, testing, and part replacement.
testing
We fully load our equipment with various utilities and software. Rest assured, it’ll work for a very long time.
Power supply
The components work at full capacity and are tested for several days. If we are not sure about something, we replace it.
of all equipment
Our regulations imply the same testing for both used and new equipment. Oh, and by the way, we complement the official manufacturer’s warranty with our own warranty.