In a significant development for the artificial intelligence sector, the Chinese AI startup DeepSeek has utilized Huawei's Ascend chips for the post-training of its DeepSeek V4 Pro model. This marks a notable departure from traditional reliance on Nvidia and AMD processors for AI training, highlighting advancements in China's chip manufacturing capabilities for AI technologies.
Researchers employed a computing cluster featuring approximately 1,000 Huawei Ascend 910C chips to conduct the post-training process. This involved a full parameter training approach, enabling the model to be completely updated without any structural changes. The collaboration included prominent institutions such as Shenzhen Loop Area, Shenzhen Campus of Harbin Institute of Technology, and the Shenzhen Institute of Big Data.
The distinction between output generation and post-training is crucial in the realm of large language models (LLMs). While output generation allows a pre-trained model to respond to user queries, post-training focuses on refining the model's ability to understand and execute human commands, safety protocols, and various operations. This advancement is expected to enhance the self-sufficiency of China's AI industry.
Previously, training for AI models like DeepSeek V3 relied on Nvidia chips, specifically a cluster of 2,048 Nvidia H800 processors, which are currently restricted in access. As DeepSeek prepares for a significant funding round aimed at raising approximately 50 billion yuan (around 7.4 billion dollars), it positions itself as a formidable competitor to established models like ChatGPT, offering powerful open-source LLMs with lower training costs.
This shift towards using Huawei chips not only reflects China's growing capabilities in AI technology but also presents a competitive challenge for companies that have long dominated the AI training market.
Informational material. 18+.
