China AI Chip Landscape 2025: Ascend, Cambricon, Hygon — Who Will Dominate?
Escalating U.S. export controls are forcing China's AI chip industry to accelerate self-reliance. By 2025, the discussion around domestic Chinese AI chips has shifted from "are they usable?" to "which one should I choose?"
This article systematically reviews the major players, core products, and actual deployment status of domestic AI chips, helping developers and procurement decision-makers understand the competitive landscape.
Tier 1: Huawei Ascend
Products: Ascend 910B (training), Ascend 310P/310 (inference)
Architecture: Da Vinci — 3D Cube matrix compute units
Core Specifications:
| Metric | Ascend 910B | Ascend 310P | Ascend 310 |
|---|---|---|---|
| FP16 compute | 400 TFLOPS | — | — |
| INT8 compute | 640 TOPS | 70 TOPS | 22 TOPS |
| Memory | 64GB HBM2e | 24GB LPDDR4X | 8GB LPDDR4 |
| TDP | 310W | 75W | 8W |
| Process | 7nm | 12nm | 12nm |
Ecosystem Status:
- CANN software stack: analogous to CUDA, a complete stack from drivers to compilers
- torch_npu: PyTorch's Ascend backend, with API highly consistent with CUDA
- MindSpore: Huawei's indigenous framework, but with limited market acceptance
- LLM adaptation: mainstream models such as Llama, Qwen all adapted
Actual Deployment: According to public data, Ascend 910B has been deployed in 6,000+ chips within Huawei's Pangu large model cluster.
Overall Assessment: The undisputed leader among domestic AI chips. The most complete software ecosystem, highest market share in government and enterprise. Training performance approaches 60-70% of H100, with competitive inference price/performance.
Tier 2: Cambricon & Hygon
Cambricon Siyuan MLU
Products: Siyuan 590, Siyuan 370
Positioning: AI training + inference
Key Information:
- Siyuan 590 compute targets A100 (FP32 ~30 TFLOPS, INT8 ~300 TOPS)
- Proprietary MLUarch architecture + BangC programming language
- PyTorch/TensorFlow adaptation already available
- Primarily deployed in smart cities, security, research, and other fields
Current Status: Cambricon was once the most-watched AI chip unicorn, but has faced commercialization difficulties and persistent losses in recent years. Product iteration pace lags behind Ascend, with market share being squeezed.
Hygon DCU (Deep Computing Unit)
Product: ShenSuan Z100
Architecture: CUDA-compatible (based on AMD ROCm path)
Key Information:
- ShenSuan No. 1 FP32 compute ~15 TFLOPS
- Biggest selling point: compatible with CUDA API, low migration cost
- Primarily deployed in supercomputing centers, financial institutions, and other Xinchuang scenarios
- Process constrained by foundry limitations
Current Status: Hygon's compatibility path lowers software migration costs in the short term, but long-term is constrained by AMD's ecosystem development.
Tier 3: Startups and Cross-Industry Players
Enflame Tech YunSui T21
- Targeting cloud AI training
- Proprietary GCU architecture + YuSuan software stack
- PyTorch adaptation available
- Won orders from multiple telecom operators and government projects
Biren Technology BR100/BR20X
- BR100 claims FP16 compute of 1,000+ TFLOPS (theoretical peak)
- But actual deployment progress lags behind claims
- Pivoted to a more pragmatic product path after 2024
Moore Threads MTT S5000
- Full-function GPU (graphics + compute + AI)
- MUSA architecture compatible with CUDA API
- Driver and software stack maturity improving, but still some distance from production-grade AI training
- Better suited for inference and small-scale training
Baidu Kunlun P800
- Baidu's indigenous AI chip
- Deployed in internal scenarios such as Baidu Search, Intelligent Cloud, autonomous driving
- Limited public technical details, but internally validated at scale
Domestic AI Chip Cross-Comparison
| Chip | FP16 Compute (TFLOPS) | Memory (GB) | CUDA Compatible | Training Capability | Deployment Scale |
|---|---|---|---|---|---|
| Ascend 910B | 400 | 64 HBM2e | ❌ CANN | ✅ Strong | 6,000+ |
| Cambricon 590 | ~300 | — | ❌ BangC | ⚠️ | 1,000s |
| Hygon DCU Z100 | ~30 (FP32) | — | ⚠️ ROCm path | ⚠️ | 1,000s |
| Enflame T21 | ~200 | 32 HBM2e | ❌ Proprietary | ✅ | 100s |
| Biren BR100 | ~1,000 (claimed) | — | ⚠️ | ⚠️ | Limited |
| Baidu Kunlun P800 | — | — | ❌ Proprietary | ⚠️ | Internal |
| Moore Threads MTT S5000 | ~100 | 32 GDDR6 | ⚠️ MUSA | ❌ Inference-first | — |
Software Ecosystem Comparison (Key Decision Factor)
| Chip | PyTorch | vLLM Inference | Hugging Face | CUDA Code Migration Cost |
|---|---|---|---|---|
| Ascend 910B | ⚠️ torch_npu | ⚠️ Community | ⚠️ Partial | Medium (need device name change + operator adaptation) |
| Hygon DCU | ⚠️ ROCm backend | ⚠️ | ⚠️ | Low (compatible with CUDA API) |
| Cambricon 590 | ⚠️ | ❌ | ❌ | High (BangC language) |
| Enflame T21 | ⚠️ | ❌ | ❌ | High |
| Moore Threads MTT | ⚠️ | ❌ | ❌ | Medium (MUSA compatible with CUDA) |
Selection Recommendations
Government / Xinchuang Projects
Ascend 910B first choice. Reasons:
- Most complete software ecosystem, strongest community support
- Ascend + Kylin/UOS combination is the Xinchuang standard
- CANN toolchain maturity leads other domestic solutions by 2-3 years
- Huawei's technical support and documentation are the most comprehensive
CUDA Legacy Code Migration
If you don't want to rewrite large amounts of code:
- Hygon DCU (ROCm compatibility path) has the lowest migration cost
- Moore Threads MTT (MUSA compatibility path) suitable for inference scenarios
- Ascend's torch_npu has medium migration cost, but the best long-term ecosystem return
Pure Inference Scenarios
- Ascend 310P: most cost-effective domestic inference card
- Moore Threads MTT S5000: if the requirement is a domestically produced full-function GPU
- Cambricon 370: has existing strength in specific scenarios (vision, security)
2025-2026 Outlook
- Ascend 920 is coming: the next generation Ascend will use more advanced process, targeting FP8 compute to match H200
- EDA tool domestication: indigenous substitution of chip design tools will help more startups accelerate iteration
- CUDA compatibility becoming standard: all domestic chips will at least provide a CUDA API compatibility layer
- Inference market share accelerating: domestic chips will be the first to reach NVIDIA-replacement level in inference scenarios
- At-scale deployment validation: more "10,000-card cluster" domestic solutions will land in telecom and financial industries
Key judgment: Chinese domestic AI chips will transition from "usable" to "good" in 2025-2026. The training performance gap remains (1-2 generations behind), but inference scenarios already meet replacement conditions.
On MirrorFrog you can find driver downloads, development documentation, and detailed specifications for all the above domestic chips.