Huawei Atlas 900 SuperPoD A2
Atlas 900 SuperPoD A2
- Cards
- 256
- Scale-up domain
- 256
- Total memory
- 32.0 TB
- Rack power
- 200 kW
- Scale-up
- HCCS-v2
- Cooling
- liquid
CHINA HUB
Chinese AI inference hardware
Coverage: 9 vendors · 14 accelerators · 2 super-pod blueprints (incl. CloudMatrix 384). All numbers carry evidence citations; missing fields are honestly flagged.
Chinese silicon × frontier models matrix
🟢 measured · 🟡 vendor-claimed (unverified) · — unsupported / unknown. Click a cell for detail.
| HW ╲ Model | DeepSeek V4 Flash | DeepSeek V4 Pro | Kimi K2.6 | MiniMax M2.7 | GLM-5.1 | Qwen3.6 Plus | Mistral Small 4 | GLM-5 Reasoning | Qwen3.5 397B Reasoning | Gemma 4 26B | Mistral Large 3 | GPT-OSS | Llama 4 Maverick | Llama 4 Scout |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| 壁仞 BR100 | — | — | — | — | — | — | — | — | — | — | — | — | — | — |
| 壁仞 BR104 | — | — | — | — | 🟢 | — | — | — | — | — | — | — | — | — |
| 寒武纪 MLU370-X8 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 |
| 寒武纪 思元 590 | 🟡 | 🟡 | 🟢 | 🟡 | 🟡 | 🟢 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 |
| 燧原 云燧 T21 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 |
| 昇腾 910B | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 |
| 昇腾 910C | 🟡 | 🟢 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 |
| 昇腾 950 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 |
| 海光 DCU K100 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟢 |
| 海光 DCU Z100 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 |
| 天数智芯 天垓 100 | — | — | — | — | — | — | — | — | — | — | — | — | — | — |
| 沐曦 曦云 C500 | — | — | — | — | — | — | — | — | — | 🟢 | — | — | — | — |
| 摩尔线程 MTT S4000 | 🟢 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 |
| 平头哥 含光 800 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 |
Generation genealogy
Year-by-year evolution per vendor. Click a node for detail.
| Vendor | 2019 | 2020 | 2021 | 2022 | 2023 | 2024 | 2025 | 2026 |
|---|---|---|---|---|---|---|---|---|
| Biren Technology | · | · | · | BR100 | BR104 | · | · | · |
| Cambricon | · | · | · | MLU370-X8 | 590 | · | · | · |
| Enflame | · | · | · | · | T21 | · | · | · |
| Huawei Ascend | · | · | · | · | 910B | 910C | · | 950 |
| Hygon | · | · | · | Z100 | · | K100 | · | · |
| Iluvatar CoreX | · | · | · | · | 100 | · | · | · |
| MetaX | · | · | · | · | · | C500 | · | · |
| Moore Threads | · | · | · | · | · | S4000 | · | · |
| T-Head (Pingtouge) | 800 | · | · | · | · | · | · | · |
Software ecosystem comparison
Programming model · operator library · inference engine · model zoo.
| Vendor | Programming model | Operator library | Inference engines | Model zoo |
|---|---|---|---|---|
| Biren Technology | BIRENSUPA | BIRENSUPA op library | — | — |
| Cambricon | BANG / Neuware | CNNL | LMDeploy | — |
| Enflame | TopsRider | Enflame SDK | vLLM | — |
| Huawei Ascend | CANN / Ascend C | AscendCL | MindIEvLLMLMDeploy | link ↗ |
| Hygon | DTK / HIP | DCU op library | vLLM | — |
| Iluvatar CoreX | IxRT / CoreX | CoreX op library | — | — |
| MetaX | MACA / MetaX SDK | MetaX op library | — | — |
| Moore Threads | MUSA | MUSA Toolkit | vLLM | — |
| T-Head (Pingtouge) | HanGuangAI | HanGuang op library | HanGuangAI | — |
Chinese super-pods
Rack-scale scale-up systems benchmarked against NVIDIA NVL72.
Huawei CloudMatrix 384
昇腾超节点 CloudMatrix 384
- Cards
- 384
- Scale-up domain
- 384
- Total memory
- 48.0 TB
- Rack power
- 600 kW
- Scale-up
- lingqu
- Cooling
- liquid