Huawei Engineers On-Site at DeepSeek, DeepSeek Plans to Use Ascend 950DT for Pre-Training as Early as Year-End
Sources familiar with the matter reveal that Huawei engineers have been stationed long-term at DeepSeek's offices, integrating large model training requirements into the design of next-generation chips. DeepSeek plans to conduct large-scale pre-training using Huawei's Ascend 950DT as early as the end of this year. A single full pre-training run may require over 10,000 chips operating continuously for several weeks, with the exact timeline dependent on testing progress. Huawei plans to allocate the majority of the supply of the Ascend 960DT, scheduled for launch in Q1 2027, to DeepSeek. DeepSeek open-sourced the TileLang toolkit adapted for the Ascend 950 on September 30, while continuing to seek additional NVIDIA chips.
Source: PANewsRead original
- #AI