NPU Accelerator
Atlas 300i Duo 96GB — Single Card
Dual Ascend 310B NPU · 96GB LPDDR4X ECC · 280 TOPS INT8 · 150W TDP. Run DeepSeek V4-Flash, Qwen 72B and Llama 3 70B fully locally, privately, without rate limits or subscriptions.
€2,150
We are unable to ship to your location.
Technical Specifications
| Total memory | 96GB LPDDR4X ECC (2× 48GB) |
| Architecture | Dual Ascend 310B NPU |
| Compute | 280 TOPS INT8 · 140 TFLOPS FP16 |
| Interface | PCIe Gen4 x16 |
| Power | 150W TDP |
| Virtualisation | Up to 7 virtual NPUs per processor |
| Models supported | 70B–100B+ parameter at reasonable quantisation |
| Condition | Brand new and unused |
Software Ecosystem
This card uses Huawei's CANN / Ascend stack — not NVIDIA CUDA. Setup requires the Ascend toolchain. Supported paths include llama.cpp (CANN backend) and PyTorch via torch-npu.
Not plug-and-play, but well-documented for technically competent buyers. DeepSeek V4 was co-developed with Huawei specifically for the Ascend platform.
Buyer is responsible for verifying software compatibility before purchase. The CANN / Ascend ecosystem is not CUDA-compatible.
Huawei Official Documentation
ServerFlow — Huawei Taishan 2280 V2 Review (translated)
Compatible Platforms
- ✓ Taishan 200 2280 V2 (ARM) — recommended
- ✓ Atlas 800 3000 — supported
- ✓ Huawei RH2288H V5 (x86) — requires driver recompile