I. Ascend 910B Series: Four Official Models and Application Scenarios
Based on Huawei's official specifications and deployment solutions:
| Model | FP16 Compute (TFLOPS) | VRAM | Server Model | Key Applications |
|---|
| 910B4 | 280 | 32GB HBM2 | Atlas 800I A2 | Edge inference, industrial inspection |
| 910B3 | 313 | 64GB HBM2 | Atlas 800T A2 | General training/inference, data centers |
| 910B2 | 376 | 64GB HBM2 | Atlas 800T A2 | High-precision R&D, financial modeling |
| 910B1 | 414 | 64GB HBM2 | HuaKun AT900 A2 (Liquid-cooled) | HPC, billion-parameter model training |
Critical Notes:
B4's 32GB VRAM cannot support training billion-parameter models (e.g., DeepSeek R1 requires ≥64GB)
B1 enables 8-card/4U high-density deployment (PUE≤1.15)
Beware misleading marketing positioning B4 as "training card" – its compute is only 67% of B1
II. Ascend 910C: Dual-Die Breakthrough & Production Challenges
The 910B successor achieves performance leap through system-level innovation:
Performance: Dual-die design delivers 780-800 TFLOPS FP16, 3.2TB/s memory bandwidth (96% of NVIDIA H100)
Production: 35-36% yield at SMIC's 7nm DUV process (target: 40%), $14.5k/die vs. 910B's $9k
Cluster Efficiency: 910C-powered CloudMatrix 384-node cluster achieves 300PFlops, 45% faster training
III. User Profiles & Applications: Market Penetration
Core Users
Government/Enterprise:
Industry Leaders:
State Grid Chongqing: 19.2PFlops 910B cluster for power equipment maintenance
Wuhan Optics Valley: 3,000 910B cards for industrial AI inspection (99.2% accuracy)
Global Expansion: Saudi Supercomputer Center replacing 40% NVIDIA orders with Ascend
Performance Highlights
LLM Training: 910C trains trillion-parameter MoE models 30% faster
Fintech: Shanghai Songjiang's fraud detection model: <5ms latency
Edge: B4 enables real-time inspection (≤10ms response)
IV. Compute Rental Market: Policy Incentives & Innovation
Economic Advantages
Pricing: 8-card 910B server: $4.7k/month (38% of H100 equivalent)
Subsidies: Whitelisted data centers enjoy state subsidies (e.g., Xi'an 15-20% cheaper)
Service Innovations
Tiered Pricing (Shanghai): SMBs rent 10-card clusters with 1st-year deposit waiver
Compute+Algorithm Bundles (Shenzhen): Pre-installed optimized PyTorch cuts dev cycle by 30%
Cross-Region Scheduling (GBA Hub): ≤3ms latency for "East Data West Computing" projects
V. National Deployment: 8 Major Hubs (As of July 2025)
| City | Key Node | Deployment | Primary Use Cases |
|---|
| Beijing | Zhongguancun | 70% 910B | LLM R&D, biomedicine |
| Shanghai | Songjiang | 5,000+ 910B | Fintech, autonomous driving |
| Shenzhen | Peng Cheng Lab | 384×910C (Phase 1) | "Pangu" LLM training |
| Wuhan | Optics Valley | 3,000×910B | Industrial simulation |
| Chongqing | Provincial Hub | 6×Atlas 910B | Smart grid maintenance |
| GBA | China Telecom Hub | 384×910C | Trillion-parameter MoE training |
Secondary Nodes: Hangzhou ZhiJiang Lab, Chengdu Hub, Nanjing Jiangbei District
Conclusion: Progress and Challenges
Ascend 910B/C drives China's AI compute sovereignty through:
✅ Scenario Coverage: B4-B1 for edge-to-HPC, 910C clusters rivaling global leaders
✅ Rental Market Boom: Cost advantage + policy support (e.g., China Telecom IDC revenue ↑56.9%)
⚠️ Challenges: 910C yield (35%), ecosystem maturity, FP8 support (910D expected 2026)
--The above article is purely AI-generated. If there is any infringement, please contact Cloud Engine/Cloud GPU Rental Platform for deletion!