Huawei Ascend 920: China's Highest Bandwidth at 4 Tbps + 3× H20 Compute for Domestic Substitution
Huawei Ascend 920 (昇腾 920) entered large-scale mass production in 2025 H2, representing a major breakthrough for Chinese domestic AI chips. This article analyzes its specifications, comparison with NVIDIA H20, the CloudMatrix 384 Ultra system, and its significance for China's AI industry.
Core Specifications
| Item | Ascend 910C | Ascend 920 | Improvement |
|---|---|---|---|
| Architecture | Da Vinci v3 | Da Vinci v4 | New generation |
| Process | 7nm | 6nm (SMIC domestic) | More advanced |
| Chiplets | 2× (dual die) | 2× | same |
| HBM capacity | ~128 GB | ~96 GB | slight decrease |
| HBM bandwidth | 3.2 Tbps | 4 Tbps | 1.25× |
| BF16 compute | 780 TFLOPS | 900+ TFLOPS | 1.15× |
| FP16 compute | 1,560 TFLOPS | 1,800 TFLOPS | 1.15× |
| INT8 compute | 3,120 TOPS | 3,600 TOPS | 1.15× |
| TDP | ~310 W | ~400 W | +29% |
| Release date | 2025-04 | 2025 H2 | — |
4 Tbps bandwidth = China's highest domestic HBM bandwidth, a 25% improvement over Ascend 910C. The 900+ BF16 TFLOPS compute also surpasses 910C.
Ascend 920 vs NVIDIA H20 (Target Comparison)
NVIDIA H20 is the "compliance" AI chip specifically designed for the Chinese market under U.S. export controls:
| Metric | Ascend 920 | NVIDIA H20 |
|---|---|---|
| Positioning | Domestic substitution | China-compliant AI chip |
| Process | 6nm (SMIC) | TSMC 4N (partially domestic after restrictions) |
| Memory | ~96 GB | 96 GB HBM3 |
| Memory bandwidth | 4 Tbps | 4.0 Tbps |
| BF16 compute | 900 TFLOPS | 296 TFLOPS |
| BF16 compute ratio | 3× | 1× (baseline) |
| Interconnect | HCCS 1.2 Tbps | NVLink 900 GB/s |
| Software | CANN + MindSpore | CUDA (restricted) |
| Import compliance | ✅ Domestic | ⚠️ U.S. export controls |
💡 Ascend 920 significantly leads H20 in BF16 compute (3×), with 4 Tbps bandwidth on par with H20. This is a key victory for domestic substitution.
CloudMatrix 384 Ultra System
Ascend 920 will be used in the CloudMatrix 384 Ultra supernode system:
| Item | Configuration |
|---|---|
| Chip count | 384 Ascend 920 chips |
| Rack count | 16 (12 compute + 4 network) |
| Total HBM | ~36 TB (96GB × 384) |
| Interconnect | Fully optical mesh, 8,000+ LPO optical modules |
| BF16 compute (system) | ~345 PFLOPS (estimated 900 × 384) |
| TDP (system) | ~150 kW |
CloudMatrix 384 Ultra system-level BF16 compute of ~345 PFLOPS ≈ 2.4× NVIDIA GB200 NVL72 cluster (~144 PF FP8 dense).
Why Ascend 920 Is the Key Victory for Domestic Substitution?
1. First Time Surpassing H20 by 3× in Compute
| Period | Domestic | NVIDIA China Edition | Multiple |
|---|---|---|---|
| 2023 | 910B = 320 TFLOPS | H20 = 296 TFLOPS | 1.08× |
| 2024 | 910B = 320 TFLOPS | H20 = 296 TFLOPS | 1.08× |
| 2025 H1 | 910C = 780 TFLOPS | H20 = 296 TFLOPS | 2.6× |
| 2025 H2 | 920 = 900 TFLOPS | H20 = 296 TFLOPS | 3.0× |
Starting from 2025 H2, Chinese domestic AI chip compute stably surpasses H20 by three times.
2. SMIC 6nm Domestic Process
Ascend 920 uses SMIC N+1 / N+2 6nm process:
- ✅ Fully indigenous and controllable
- ✅ Not subject to U.S. export controls
- ⚠️ Yield and cost still lag behind TSMC 4N
3. 4 Tbps — China's Highest Domestic HBM
Ascend 920's 4 Tbps HBM bandwidth:
- First domestic chip to reach 4 Tbps level (previous max 3.2 Tbps)
- On par with H20
- Presumed to use CXMT (ChangXin Memory Technologies) HBM3 or indigenous HBM
4. CANN + MindSpore Software Stack
- CANN 8.x (Compute Architecture for Neural Networks): analog to CUDA
- MindSpore 2.4+: Huawei's indigenous AI framework
- PyTorch 2.3+ MindSpore backend: PyTorch compatible
- vLLM 0.7+ Ascend backend: low-latency inference
- ONNX-Runtime Ascend backend: cross-framework inference
- Atlas 900/950 series servers: OEM complete systems
China Market Deployment Status
Scaled-Up Customers
| Customer | Application |
|---|---|
| China Mobile | Large model training (990M customers) |
| China Telecom | Intelligent customer service + business insights |
| China Unicom | Government + industry AI |
| State Grid | Power grid scheduling + fault prediction |
| CNPC | Exploration + logistics optimization |
| Major banks | Risk control + anti-fraud |
| Internet companies (Baidu, Alibaba, Tencent) | LLM inference |
Industry Layout
- Government: 100% domestic requirement
- Finance: policy-driven domestic requirement
- Telecom: fast HBM domestication progress
- Energy: fast HBM domestication progress
- Internet: sensitive workloads shifting to domestic
- Education / Healthcare: gradual domestication
Limitations and Challenges
| Limitation | Impact |
|---|---|
| FP8/FP4 support | Ascend 920 still BF16/FP16-primary, FP8 optimization in progress |
| HBM capacity | 96 GB is below NVIDIA Rubin R200 288 GB / AMD MI400 432 GB |
| CUDA compatibility | CANN 8 still requires migration; direct CUDA app execution is limited |
| SMIC 6nm yield | 10-20% lower yield than TSMC 4N |
| HBM source | CXMT HBM production capacity limited |
| Interconnect bandwidth | HCCS 1.2 Tbps far below NVLink 6 (3.5 TB/s) |
Comparison with Contemporaneous Domestic Chips
| Vendor | Chip | BF16 Compute | HBM Bandwidth | Mass Production |
|---|---|---|---|---|
| Huawei | Ascend 920 | 900 TFLOPS | 4 Tbps | 2025 H2 |
| Huawei | Ascend 910C | 780 TFLOPS | 3.2 Tbps | 2025-04 |
| Cambricon | Siyuan 590 | ~480 TFLOPS | 2.4 Tbps | 2024 |
| Moore Threads | MTT S5000 | ~250 TFLOPS | 1.6 Tbps | 2024 |
| Biren | BR104 | ~300 TFLOPS | 1.6 Tbps | 2024 |
| Iluvatar | CoreX Bi-150 | ~200 TFLOPS | 1.2 Tbps | 2024 |
Huawei Ascend 920 maintains a clear lead among Chinese domestic AI chips.
Detailed Product Pages
- Huawei Ascend 920 Full Specs
- Huawei Ascend 910C (Previous Gen)
- Huawei Ascend 910B (First Mass Production Gen)
- Huawei Ascend 910D (High-End)
- NVIDIA H100 NVL (Target)
- AMD MI300X (Overseas Alternative)
- Future Roadmap
Summary
Huawei Ascend 920 is a key victory for Chinese AI chips in 2025 H2:
- 900+ BF16 TFLOPS = 3× H20 — first time stably surpassing H20 by three times
- SMIC 6nm domestic — indigenous and controllable
- 4 Tbps — China's highest domestic HBM bandwidth — HBM domestication breakthrough
- CloudMatrix 384 Ultra system — single system surpasses GB200 NVL72
- CANN + MindSpore — maturing software ecosystem
Starting from 2025 H2, China's AI industry enters a new phase where "domestic chips can independently support large-scale AI applications."