DeepSeek has released programming infrastructure for Huawei’s Ascend chips, including an open-source language aimed at simplifying AI accelerator development.
DeepSeek has partnered with Huawei to develop software infrastructure for Ascend AI chips, while open-sourcing components that support programming and distributed computing on the platform. The release includes compute and communication libraries, along with an Ascend version of TileLang, a high-level open-source programming language designed for AI accelerator development.
TileLang provides developers with a higher-level programming model while allowing code to target the underlying hardware. DeepSeek says the language can simplify programming compared with lower-level approaches such as Nvidia’s CUDA, while still allowing developers to optimise workloads for the hardware. The Ascend version wraps the platform’s low-level Ascend C instructions, and DeepSeek says its training operators now have high-performance Ascend implementations.
The open-source release also covers computing and distributed communication libraries for Huawei’s Ascend platform. DeepSeek said the components correspond to versions it has previously released for Nvidia hardware, giving developers a software path for adapting AI workloads to Ascend-based systems. This adds to the growing open-source software ecosystem around Huawei’s AI processors.
DeepSeek and Huawei have also worked on a supernode architecture using 128 Ascend 950 chips, with optimisation across both computation and communication. Huawei has separately been expanding the open-source development of its Ascend software stack, including the Compute Architecture for Neural Networks (CANN), which supports projects such as PyTorch, Triton and vLLM.
The release gives developers access to more open-source infrastructure for building and optimising AI workloads beyond Nvidia’s CUDA ecosystem. Rather than replacing CUDA directly, the combination of TileLang and Ascend software provides another programming stack for AI accelerators and supports DeepSeek and Huawei’s efforts to develop a more open, hardware-specific software ecosystem.
