DeepSeek has released open-source software for programming Huawei’s Ascend AI accelerators, including libraries for computation and moving data between chips. Huawei supported the work, and the companies also optimized a supernode made from 128 Ascend 950 chips.
At the center is TileLang, a language originally developed by Peking University researchers. DeepSeek has used it for about a year and describes it as a simpler programming model than Nvidia’s CUDA while still exposing enough control to extract hardware performance. The company first tested TileLang on older Nvidia GPUs before making it central to its own model-development work.
The release targets the hardest part of replacing imported accelerators: software. Nvidia’s advantage includes an estimated four million CUDA developers and mature tools that coordinate many chips, not only the performance of individual processors. Domestic Chinese models have advanced faster than local hardware and its supporting ecosystem.
Huawei recently announced new processors and supernode systems for wider model-training use next year, while acknowledging that it cannot satisfy all domestic demand. Open-source tools can make Ascend hardware easier to adopt, but a language and several libraries do not instantly reproduce CUDA’s years of documentation, integrations and developer experience. Real progress will depend on model support, debugging tools and performance across varied workloads.