DeepSeek has released six open-source programming tools for Huawei’s Ascend chips, covering computation and communication between processors. The company announced the release on September 30 through its WeChat account.
Nvidia’s CUDA libraries and tools help keep AI developers on its hardware. US export controls have pushed Chinese firms to seek alternatives, but competing chips also need software that developers can use efficiently.
The six components correspond to tools DeepSeek previously released for Nvidia hardware. With Huawei’s support, the companies also jointly optimized computation and communication in a system linking 128 Ascend 950 chips.
TileLang, a language for writing the low-level operations used in AI models, now supports Ascend 950 with native code generation and automatic scheduling and synchronization. DeepSeek says every TileLang operation it uses in training has a high-performance implementation on Ascend.
The release covers more than the language: DeepGEMM handles matrix calculations, DeepEP moves data between chips, and FlashMLA accelerates sparse attention for processing long contexts. DeepSeek says several key tests approach the hardware’s performance limits.