DeepSeek releases computation and communication tools for Huawei Ascend chips

DeepSeek has released open-source programming tools developed with Huawei for Ascend AI chips, including computation and chip-to-chip communication libraries and native Ascend 950 support in TileLang. Reuters reported the release on September 30. DeepGEMM-Ascend performs matrix calculations using the same programming interfaces as DeepSeek's existing DeepGEMM library. DeepEP-Ascend routes data between model experts and combines their outputs during training and inference. Both libraries were developed and tested on Ascend 950 hardware. The companies also optimized computation and communication on a system containing 128 Ascend 950 chips. TileLang continues to support Nvidia GPUs, while the Ascend tools rely on Huawei's underlying CANN software platform.