Specifications
System Type
4U rack-mounted AI training server
AI Processors
8x Huawei Ascend 910B NPU (64GB HBM2e each)
Aggregate NPU Memory
512GB HBM2e
Host CPUs
4x Kunpeng 920 (up to 64 cores each)
Host Memory
Up to 32x DDR4 DIMM slots, RDIMM, ECC, up to 512GB
AI Performance
Up to 3.0 PFLOPS FP16 / 6.0 POPS INT8
Accelerator Interconnect
Huawei HCCS high-speed interconnect
Networking
8x 200GE QSFP ports with RoCE
Host I/O
Up to 3x PCIe 4.0 expansion slots
Storage
8x 2.5-inch SAS/SATA bays plus NVMe options (8SFF+2NVMe config)
Cooling
Air cooling, 8x hot-swap fan modules with N+1 redundancy
Max System Power
Up to 10.4kW
Software Stack
Huawei CANN, MindSpore, MindSpeed, AscendCL, PyTorch Adapter (torch-npu), vLLM-Ascend
Supported Precisions
FP32, FP16, BF16, INT8
Virtualization
Huawei vNPU partitioning, device passthrough, container deployment
Certifications
CE, FCC, RoHS
Overview
The Huawei Atlas 800T A2 is Huawei's mainstream 4U AI training server. It integrates eight Ascend 910B NPUs — each with 64GB of HBM2e — and four Kunpeng 920 host processors, yielding 512GB of aggregated accelerator memory and up to 3.0 PFLOPS FP16 / 6.0 POPS INT8 across the eight-NPU HCCS interconnect.
Eight 200GE QSFP ports with RoCE give the platform the network bandwidth required for distributed multi-node training, while up to 32 DDR4 DIMM slots and hot-swap storage bays provide the host-side capacity for large datasets. Air cooling with N+1 redundant fans keeps deployment straightforward in standard racks.
The system runs the Huawei Ascend ecosystem — CANN, MindSpore, MindSpeed, the PyTorch Ascend adapter, vLLM-Ascend and lmdeploy — and is widely used for LLM pre-training and fine-tuning, computer vision and scientific AI. QS Compute supplies the Atlas 800T A2 for AI training clusters and high-throughput inference deployments.
Key Benefits
8x Ascend 910B with 512GB aggregate HBM2e. 3.0 PFLOPS FP16 for large-model training. 8x 200GE RoCE for distributed training scale-out. Full CANN / MindSpore stack with vLLM-Ascend support.
Applications
Large language model pre-training and fine-tuning, distributed deep-learning training, high-throughput inference and RAG, computer vision and scientific research.
Request a Quote — HUAWEI ATLAS 800T A2 — 8X ASCEND 910B AI TRAINING SERVER
QS Compute — global B2B supply of AI computing hardware, edge AI systems and accelerators. Volume pricing, 15-day sample lead time.
Get Your Quote →