It is reported that Huawei's AI semiconductor Ascend 910C has been extensively utilized in China's DeepSeek's artificial intelligence (AI) model, with the Ascend 910C achieving performance levels around 60% that of NVIDIA's AI accelerator H100. NVIDIA's H100 is the top model of the Hopper series that was released prior to the recently launched Blackwell series. So far, Huawei has brought the previous generation 910B of the 910C to market and is believed to have validated the performance before the official launch of the 910C.
According to industry sources on the 7th, Huawei's Ascend 910C has been deployed to enhance the inference capabilities of DeepSeek's AI model 'R1'. The process of training AI models with large amounts of data utilizes NVIDIA's AI accelerator H80, which is designed for the Chinese market, while the Ascend series from Huawei is utilized to improve inference capabilities.
Huawei's Ascend series is symbolic of China's semiconductor ambitions. It is a product designed for AI models that the Chinese government is heavily investing in and nurturing. Due to U.S. government regulations preventing Huawei from utilizing advanced processes from Taiwan's TSMC, production is taking place at China's largest foundry corporation, SMIC. Both Huawei's Ascend 910B and 910C are reported to have been manufactured through SMIC's 7nm (nanometer) process. While global big tech corporations like NVIDIA, AMD, and Intel utilize processes below 4nm for their AI accelerators, SMIC has been unable to import extreme ultraviolet (EUV) lithography equipment essential for processes below 7nm due to U.S. regulations, resulting in its technology being limited to 7nm.
With the global AI craze, DeepSeek has unveiled highly sophisticated AI models, drawing attention to the AI semiconductors that power them. There were analyses suggesting that DeepSeek likely utilized Huawei's semiconductors as the import of NVIDIA's high-performance semiconductors was impossible due to U.S. regulations.
China is building an independent semiconductor ecosystem centered around Huawei; however, it is still analyzed that there is a significantly high dependence on NVIDIA. Improving the inference performance of AI is possible through Huawei's AI accelerators, but NVIDIA's high-performance semiconductors remain essential for training large amounts of data. Furthermore, during the process of training AI models using NVIDIA's semiconductors, there is also a high utilization of NVIDIA's proprietary software platform 'CUDA'. DeepSeek also noted that 'the Chinese semiconductor industry has become accustomed to NVIDIA's CUDA ecosystem, established over 20 years,' adding that 'Huawei also needs to improve software along with hardware for AI data training.'
A semiconductor industry official said, 'It is encouraging that although it does not reach the capabilities of NVIDIA's AI accelerators amidst ongoing U.S. regulations, a semiconductor capable of developing AI models has been developed.' However, the official also remarked that 'since AI requires the establishment of both hardware and software ecosystems, achieving complete independence from NVIDIA is essentially impossible.'