DeepSeek, an artificial intelligence laboratory in China, is gaining worldwide attention as their AI model R1 is being rated on par with, and sometimes surpassing, leading US platforms like OpenAI's o1 and Meta's Llama. According to Scale AI's CEO Alexandr Wang, DeepSeek currently owns a significant number of NVIDIA GPUs, which they use to train this advanced AI model.
Tens of Thousands of NVIDIA GPUs in DeepSeek's Possession
In a recent CNBC interview, Alexandr Wang shared details about "Humanity's Last Exam" - considered the world's most challenging AI test, featuring questions from experts in mathematics, physics, chemistry, and biology. The results showed that DeepSeek R1 is a leading AI model, performing equal to or better than the best US models.
Wang noted that while the US has long been the leader in artificial intelligence, the emergence of R1 from China has changed the situation. He pointed out that DeepSeek's decision to release this model during Christmas was symbolic, as most Western countries were enjoying their holiday break.
One of the key factors behind DeepSeek's outstanding performance is their access to a large number of NVIDIA GPUs. Alexandr Wang revealed that DeepSeek currently owns approximately 50,000 NVIDIA H100 GPUs - an impressive number, especially considering that the US has banned the export of high-end GPUs like the H100 and H200 to China since 2022.
NVIDIA had previously developed alternative versions like the H800 and A800 to meet export regulations, but these versions were also banned in 2023. However, DeepSeek still has more GPUs than what the public might think, although they cannot openly discuss this due to US export control rules.
Despite having an impressive number of GPUs, Wang believes that it will become increasingly difficult for China to get more advanced GPUs because of current and future export bans. This creates major limitations for Chinese AI labs wanting to increase their computing power.
Nevertheless, looking at what DeepSeek has achieved, it's clear they are becoming a strong competitor, capable of challenging top US tech companies. The effective ownership and use of NVIDIA GPUs is helping China establish its position in the global AI race. Their success shows that even with trade restrictions, Chinese companies can still make significant progress in artificial intelligence development. The situation highlights how complex the international AI competition has become, especially regarding access to essential computing resources like high-performance GPUs.
Currently, DeepSeek is not just limited to the R1 model. The startup has developed additional models like DeepSeek-V3 and most recently DeepSeek-R1, which was launched on January 20th. DeepSeek-V3 has 175 billion parameters and was trained on 570GB of text data from various sources including books, articles, and Wikipedia.
What's remarkable is that DeepSeek-V3 only needed 2,000 NVIDIA GPUs to process the training data, compared to ChatGPT's requirement of 10,000 GPUs. This demonstrates DeepSeek's superior algorithm optimization capabilities, especially considering the restrictions caused by US AI chip export bans.
DeepSeek-R1 takes things further by implementing the "chain of thought" method, allowing the model to review and evaluate its own logic. This approach helps R1 solve more complex problems with higher accuracy, even surpassing leading models like OpenAI's GPT-4 and Anthropic's Claude Sonnet 3.5.
-----
Source: genk.vn, 27/Jan/2025, https://genk.vn/khien-ca-thung-lung-silicon-chao-dao-startup-ai-trung-quoc-deepseek-dang-so-huu-bao-nhieu-gpu-nvidia-khong-nhieu-nhu-openai-nhung-hieu-suat-lai-ngang-ngua-2025012801532808.chn
0 Comments