NPU Accelerator

Atlas 300i Duo 96GB — Quad Bundle

4× Atlas 300i Duo cards: 384GB combined LPDDR4X ECC · 1,120 TOPS INT8. Full-scale private inference cluster in a single server chassis.

€0

We are unable to ship to your location.

Technical Specifications

Total memory 384GB LPDDR4X ECC (8× 48GB)
Architecture 8× Ascend 310B NPU
Compute 1,120 TOPS INT8 · 560 TFLOPS FP16
Interface 4× PCIe Gen4 x16
Power 4× 150W TDP
Condition Brand new and unused

Software ecosystem

The Atlas 300i Duo runs on Huawei's CANN / Ascend stack. Two actively maintained inference paths work today:

  • llama.cpp (CANN backend) — widely tested for GGUF-format LLM inference
  • PyTorch via torch-npu — full Python ML stack; supports training and fine-tuning
Not CUDA-compatible. Buyer is responsible for verifying software compatibility before purchase. The CANN / Ascend ecosystem requires separate setup from NVIDIA pipelines.

What runs on it

DeepSeek V4 was co-developed with Huawei specifically for the Ascend platform and runs natively. Other validated models include:

  • DeepSeek V4-Flash, V3, and R1
  • Qwen 2.5 72B
  • Llama 3 70B / Llama 3.1 70B
  • Mistral 7B / Mixtral 8×7B
  • ChatGLM3-6B and other Chinese-language models

Atlas 300i Duo 96GB — Quad Bundle — product demonstration

Compatible Platforms

  • ✓ Taishan 200 2280 V2 (ARM) — recommended
  • ✓ Atlas 800 3000 — supported
  • ✓ Huawei RH2288H V5 (x86) — requires driver recompile