DeepSeek Joins Huawei Chip Design, Plans Ascend 950DT Pretraining

DeepSeek is participating in the technical development of Huawei's next-generation AI chips from the design stage, and plans to use Huawei's latest flagship AI chip, the Ascend 950DT, for pretraining its models as early as the end of this year, according to aitimes.com. Testing is still ongoing, so the introduction schedule could change.

Huawei has dispatched engineers to work long-term at DeepSeek's offices to incorporate chip design requirements. The latest test used more than 1,000 processors.

Huawei plans to allocate most of the next-generation training chip Ascend 960DT to DeepSeek rather than distributing supply among other Chinese AI developers. The Ascend 960DT is scheduled for release next year, with the Ascend 970 targeted for 2028 and the Ascend 980 for 2029.

DeepSeek last month released a version of TileLang that runs on Huawei's Ascend 950 chip along with five sets of compute and data-transfer code. Huawei has reportedly begun mass shipments of SuperNode, a very large computing system equipped with the Ascend 950DT.

Performance gap and US restrictions

In 2020, the US restricted Huawei from using Taiwan's TSMC, and has since limited China's access to advanced high-bandwidth memory (HBM). According to Epoch AI, Nvidia's B300 chip has about 7 times the basic arithmetic performance of Huawei's Ascend 950, and Huawei has projected it will lag Nvidia by 3-4 years in single-chip performance through 2030.