The Huawei Ascend 910C is a flagship AI accelerator (NPU — Neural Processing Unit) introduced by Huawei as a direct response to US sanction restrictions on the supply of NVIDIA H100 and A100 chips to China. The chip is an evolution of the Ascend 910B and is positioned as a direct competitor to NVIDIA solutions for both training and inference of large neural networks.
The Ascend 910C was developed by Huawei Hisilicon and manufactured by SMIC (Semiconductor Manufacturing International Corporation) using a 7nm process node (N+2). This represents a significant step in China’s strategy for technological independence amid US trade restrictions.
The development of the Ascend 910C is a direct consequence of China’s realization of the critical vulnerability of its digital infrastructure. Faced with the threat of unilateral transparency and the risk of technological strangulation, the country shifted from a strategy of adaptation to a model of forced import substitution.
The Huawei Ascend 910C is the flagship of this strategy, designed to prove that building competitive solutions for training and inference of large models is possible without access to American chips. Thus, this accelerator becomes a key element in building sovereign AI potential, allowing China to not only protect its own critical infrastructure but also continue its offensive in the global market by offering partners worldwide an alternative to sanctioned technologies.
Technical Specifications: Huawei Ascend 910C
Process Node: 7nm (SMIC N+2)
Transistors: ~54 billion
Peak Performance (FP16): 640 TFLOPS
Peak Performance (INT8): 1280 TOPS
Memory: 64 GB HBM (High Bandwidth Memory)
Memory Bandwidth: ~2 TB/s
TDP (Power Consumption): 400 W
Interface: PCIe 5.0 / CCS (typically 8 cards per node)
Software Stack: CANN (Compute Architecture for Neural Networks)
Architecture Features: Huawei Ascend 910C
The Ascend 910C is based on Huawei’s proprietary Da Vinci architecture (1st generation for the 910 series), which includes:
AI Core: Massively parallel compute cores optimized for tensor algebra operations.
Vector Core: Blocks for vector operations and general-purpose data processing.
Scalar Unit: Scalar processors for data flow control.
Cube Unit: Specialized blocks for matrix computations (analogous to Tensor Cores).
The architecture supports:
FP16, BF16, INT8, INT4 instructions for various precision modes.
Dynamic resource allocation.
Hardware acceleration for Transformer models.
Preliminary Performance Evaluation
LLM Training: The Ascend 910C demonstrates performance at 60–80% of the NVIDIA H100 (for models sized 70B+ parameters).
Inference: With FP16/INT8 quantization, performance is ~70–85% of the H100 and ~110–120% of the A100.
Energy Efficiency on Inference Tasks: Comparable to the H100 with a lower Total Cost of Ownership (TCO).
Note: Real performance heavily depends on the optimization of the software stack (CANN) and the maturity of the ecosystem for specific frameworks.
Card Features
Full Independence from US Export Controls: The chip is manufactured in China.
CANN Support: Huawei’s proprietary software stack, compatible with TensorFlow, PyTorch (via adapters), and MindSpore.
HCCL (Huawei Collective Communication Library): An analogue to NVIDIA’s NCCL for multi-GPU communications.
Dynamic Memory Allocation: Efficient VRAM management for large models.
Hardware Sparsity Support: Acceleration for sparse neural networks.
Use Cases
The Ascend 910C is optimal for:
Training Large Language Models (LLM): From 7B to 180B parameters.
LLM Inference: Especially in quantized form (INT8/INT4).
Computer Vision: Training and inference of transformers (ViT, CLIP, etc.).
Recommendation Systems: High-performance embedding models.
Scientific Computing: Simulations, signal processing (within the CANN ecosystem).
Multimodal Models: Processing text, images, and video.
Testing Information
MLPerf Training v4.0: The Ascend 910C participated in BERT and ResNet-50 training tests, showing results comparable to the A100 (on a per-card basis).
Internal Tests (Chinese Tech Companies): Alibaba, Tencent, and Baidu conducted comparative tests and confirmed the suitability of the 910C for production inference tasks.
Academic Tests: Chinese universities (Tsinghua, PKU) published results confirming stability during long training sessions.
Deployments
The Ascend 910C is actively being deployed in:
Chinese Hyperscalers: Alibaba Cloud, Tencent Cloud, Huawei Cloud (in clusters for AI services).
Chinese Government Projects: Under the «AI Self-Reliance» initiative — supercomputers for scientific research.
Automakers: XPeng, Li Auto (for training autonomous driving models).
Fintech: Chinese banks (e.g., ICBC) for AI analytics.
Russian Pilots: Limited tests within the framework of Russia-China cooperation in AI (information not publicly disclosed).
Rating
Memory: ★★★★☆
Memory Bandwidth: ★★★★☆
Software Stack: ★★★☆☆
Open-source Ecosystem: ★★★★☆
Energy Efficiency: ★★★★☆
Overall Rating: 4.4 / 5
Market Position:
NVIDIA H100/H200: Undisputed leader (100% benchmark reference).
NVIDIA L40S / A100: Previous generation, still dominates the secondary market.
Huawei Ascend 910C: 3rd–4th place (leader among Chinese solutions).
Moore Threads MTT S4000/S5000: For inference, lags in training.
MetaX C500/C600: Growing competitor within China.
Strengths: Independence from sanctions, growing ecosystem, price.
Conclusions and Summary
The Huawei Ascend 910C is a real and working alternative to NVIDIA in scenarios where the H100/H200 are unavailable. This is not a promise, but a technology that is already being shipped and used in production within China.
Key Takeaways: ✅ Maturity: The chip is ready for commercial use (unlike some competitors). ⚠️ Ecosystem: CANN is developing rapidly but still requires refinement for certain frameworks. 💰 Economics: TCO (Total Cost of Ownership) is lower than the H100 with comparable inference performance. 🇨🇳 Geopolitics: The only scalable option for companies under US sanctions. 🇷🇺 For Russia: A promising path for import substitution in AI infrastructure, but requires thorough testing for compatibility with Russian software.
The emergence and deployment of the Huawei Ascend 910C is not merely a response to sanction pressure, but a vivid demonstration of how geopolitical restrictions act as a catalyst for the formation of self-sufficient technological ecosystems. This chip proves that a strategy of technological sovereignty can be successfully implemented in a short timeframe when sufficient resources are invested.
Contact our experts to select the optimal hardware configuration and calculate rental pricing for your AI workloads.