Overview
Huawei Ascend 910D is the next-generation product after 910C, expected to begin sampling in mid-to-late 2025. It targets NVIDIA Blackwell and the next-gen Rubin series. The 910D continues to use HBM2e (under export controls), but system-level design is further upgraded.
| Item | Status |
|---|
| Architecture | Further upgraded Da Vinci |
| HBM | Still HBM2e (2 generations behind NVIDIA) |
| Compute Target | Close to NVIDIA H100 (some metrics) |
| Status | Already superseded by 950 series (2026 H1) |
| Mass Production | H2 2025 |
| Customers | Baidu, Alibaba, Tencent and other major Chinese firms |
vs NVIDIA Comparison
| Metric | Ascend 910D (Est.) | NVIDIA H100 | NVIDIA B200 |
|---|
| HBM | HBM2e | HBM3 | HBM3e |
| Software Ecosystem | CANN + MindSpore | CUDA | CUDA |
| Training Maturity | Medium | Mature | Mature |
| Inference Performance | Near H100 | Baseline | Far ahead |
Key Challenges
- HBM 2 generations behind (HBM2e vs NVIDIA HBM3e)
- EDA tool restrictions: Dependent on domestic/third-party solutions
- CUDA ecosystem advantage: Major Chinese firms still have large CUDA codebases
- Software gap: CANN still needs time to catch up with CUDA
Use Cases
- High-end AI training in the Chinese market
- National-level AI projects
- Domains with strict localization requirements