Huawei Atlas 800I A2 Inference Server, 4U Kunpeng 920

The Huawei Atlas 800I A2 is a 4U inference server built on four Kunpeng 920 CPUs and Ascend A2 processors, designed around eight compute modules rather than a single large accelerator. Beijing Xinchuan Technology supplies and configures these systems for generative model inference, including intelligent customer service, text generation and knowledge retrieval, where throughput per rack matters more than peak training performance.

Host memory runs across 32 DDR4 slots at up to 3200 MT/s, with 16 GB, 32 GB or 64 GB modules, which is what allows large models to be staged in host memory before they are pushed to the compute modules. An optional full-interconnect build provides 392 GB/s of aggregate bandwidth inside the chassis. External networking is eight 200GE QSFP ports routed directly with RoCE, and three PCIe 4.0 slots remain free for host adapters. Four 2.6 kW power modules run in 2+2 redundancy with eight hot-swap fan modules in N+1.

Form factor: 4U AI server
CPU: 4 × Kunpeng 920
Memory: 32 × DDR4 slots, up to 3200 MT/s, 16/32/64 GB modules
Local storage: 8 × 2.5″ SATA + 2 × 2.5″ NVMe, or 4 × 2.5″ SATA + 6 × 2.5″ NVMe
RAID: 0/1/10/5/50/6/60
Internal fabric: optional full NPU interconnect, 392 GB/s
External network: 8 × 200GE QSFP direct, RoCE
PCIe: up to 3 × PCIe 4.0 slots
Power: 4 × hot-swap 2.6 kW modules, 2+2 redundant; 200–240 V AC or 240 V DC
Dimensions: 175 × 447 × 790 mm

Each system is supplied by Beijing Xinchuan Technology with configuration support and worldwide shipping. Memory population, interconnect option and drive layout are confirmed on the quotation. Contact us for current availability and a configured quote.

Parameters	Atlas 800I A2
Hardware specifications
Form factor	4U AI server
CPU	4 × Kunpeng 920
Memory	32 × DDR4 slots, up to 3200 MT/s, 16/32/64 GB modules
Local storage	8 × 2.5″ SATA + 2 × 2.5″ NVMe 4 × 2.5″ SATA + 6 × 2.5″ NVMe
RAID support	RAID 0/1/10/5/50/6/60
Internal network	Optional full NPU interconnect, 392 GB/s aggregate
External network	8 × 200GE QSFP direct, RoCE
PCIe slots	Up to 3 × PCIe 4.0
Power modules	4 × hot-swap 2.6 kW, 2+2 redundancy
Input power	200–240 V AC / 240 V DC
Cooling	Air cooled, 8 hot-swap fan modules, N+1
Operating temperature	5 °C to 35 °C
Dimensions (H × W × D)	175 × 447 × 790 mm

Related models

Frequently asked questions

What is the difference between the 800I A2 and the 800T A2?

The 800I A2 is positioned for inference and generation workloads, with an optional full-NPU interconnect at 392 GB/s and 32 DDR4 slots for large host memory. The 800T A2 is the training variant. Both use four Kunpeng 920 CPUs in a 4U chassis.

Does it support NPU interconnect?

Yes. A full-interconnect configuration is available as an option, with 392 GB/s aggregate bandwidth inside the chassis, which helps when a single model spans several NPUs.