DeepSeek, a new Chinese AI chatbot, has gained immense popularity, ranking first in the App Store across 51 countries. It is reported that DeepSeek's large-language model, the R1, has been trained using Nvidia H100 chips but utilizes Huawei's Ascend 910C for inference. This shift marks a significant advancement in making AI models more accessible and cost-efficient, with Huawei's upcoming 920C chip expected to further improve AI capabilities and challenge Nvidia's dominance in this arena.
The information comes from @Dorialexander, who points out that Ascend chips are not dealing with training, so the GPU power requirements are not that high.
DeepSeek has trained on Nvidia H800 but is running inference on the new home Chinese chips made by Huawei, the 910C.
Collection
[
|
...
]