Huawei Ascend 910C Enters Mass Production: Can Domestic AI Chips Replace NVIDIA?

Huawei's Ascend 910C AI chip has officially entered mass production.

Performance

• FP16 compute: approximately 320 TFLOPS (H100 is 400 TFLOPS)
• Power consumption: 310W
• Process node: 7nm (limited by sanctions, unable to use more advanced process nodes)
• Memory: 64GB HBM2E

Advantages

• Not subject to U.S. export controls
• Price is about 50% cheaper than NVIDIA
• Deep integration with Huawei Cloud
• Politically correct choice for domestic substitution

Disadvantages

• Weak software ecosystem (CUDA's advantage isn't the hardware—it's the developer ecosystem built over a decade)
• Can only run Huawei's own MindSpore framework
• High migration costs for many PyTorch models
• Performance ceiling limited by process node

Can It Replace NVIDIA?

Not in the short term. However, in specific scenarios (such as AI projects for government and state-owned enterprises), the Ascend 910C is a necessity—these customers must use domestic chips.

Huawei's strategy is also quite smart: instead of competing with NVIDIA on training performance for massive GPU clusters, it focuses on inference scenarios (which have lower dependency on ecosystem and lower hardware compatibility requirements).

There's still a long way to go for domestic AI chips, but the Ascend 910C at least proves this path is viable.

References:

• Huawei — Ascend