DeepSeek and Huawei released open-source software for Huawei’s Ascend AI chips on Wednesday, expanding their effort to build a Chinese computing ecosystem with less dependence on Nvidia’s hardware and programming tools.
DeepSeek said the collaboration covers high-level programming with TileLang and libraries for computation and communication. The company described its goal as an independent, self-controlled graphics-processing software ecosystem that combines ease of programming with high hardware performance.
The release includes DeepGEMM-Ascend, a version of DeepSeek’s matrix-computation library for Huawei processors. Matrix multiplication is a core operation in AI models. The repository identifies September 30 as its initial release date and lists support for Ascend 950 chips.
DeepGEMM-Ascend preserves the programming interface of the original library, allowing developers to use the same calls on supported Nvidia and Huawei platforms. It supports several numerical formats used in model computation and requires Huawei’s CANN software and a compatible PyTorch installation. The tools are intended for developers building training and inference systems on Huawei hardware.
A second library, DeepEP-Ascend, handles communication during training and inference. It includes operations that route data among the specialized components of mixture-of-experts models and combine their outputs. Its public interfaces are aligned with the Nvidia version of DeepEP.
The communication benchmarks come with an important qualification. DeepSeek measured performance using Ascend 950DT chips and a proof-of-concept Huawei hardware-development kit with additional manual configuration. The repository says that configuration is not publicly distributed and that users of earlier versions may see lower bandwidth.
Huawei has advised DeepSeek that a commercial kit for Atlas 850E, planned for mid-October, will include the configuration needed for full-bandwidth operation. The repository gives October 15 as an approximate release date, subject to Huawei’s schedule, and explicitly says the published results were not measured on that unreleased commercial version.
Some DeepEP-Ascend features remain experimental, including communication between successive stages of a model and access to remote memory. Other communication operations are still being built. The published results establish performance for the documented test setup, not an independent comparison of complete models running on Huawei and Nvidia systems.
The software release follows DeepSeek’s reported plan to acquire at least 160,000 Ascend 950DT chips for an inference facility in Inner Mongolia. That plan was reported by Bloomberg earlier this month; it was not an announcement that the chips had been delivered.
Huawei has also moved its Ascend 960DT roadmap to the first quarter of 2027. Wednesday’s libraries target the current Ascend 950 generation. Developers deploying DeepEP-Ascend are directed to use the commercial Huawei software package once it becomes available.
Sources: Reuters, DeepSeek, DeepSeek
–
By the Control Plane Editorial Team