DeepSeek Helps Shape Huawei AI Chips, Eyes Year-End Model Training
Photo: TokenPost
DeepSeek is helping shape Huawei's future AI chips around its model training needs and plans to start pretraining on Ascend hardware as early as year-end, potentially challenging Nvidia's role in China. Huawei engineers have been working inside DeepSeek's offices on designs informed by the developer's training workloads. Their work addresses how computing tasks are assigned and how data moves between chips. DeepSeek plans to use the Ascend 950DT for a pretraining attempt. Tests have involved more than 1,000 chips, but a full run could require tens of thousands. Some software needed to manage a cluster that large remains unfinished.
Source: TokenPost
