Overview
- DeepSeek announced it has partnered with Huawei to produce and open-source programming infrastructure for Huawei’s Ascend platform, including compute and communication libraries.
- The toolkit packages multiple libraries for core AI tasks such as matrix math, memory handling, kernel development, and inter-chip communication to simplify porting models to Ascend chips.
- DeepSeek and Huawei said they jointly developed a 128-chip supernode design based on Ascend 950 accelerators to optimise both computation and cross-chip communication for large training runs.
- The release promotes TileLang, a high-level open-source language that DeepSeek positions as an easier programming model than Nvidia’s CUDA while still able to reach hardware performance.
- The move is part of a wider push to build a domestic AI hardware-software stack that could lower switching costs for Chinese developers and support large-scale model training on local systems in the coming year.